Computer Methods and Programs in Biomedicine 259 (2025) 108506

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.sciencedirect.com/journal/computer-methods-
and-programs-in-biomedicine

Deep learning techniques for automated Alzheimer’s and mild cognitive

impairment disease using EEG signals: A comprehensive review of the last
decade (2013 - 2024)
Madhav Acharya a,* , Ravinesh C Deo a,*, Xiaohui Tao b,
Prabal Datta Barua a,c,d,e,f,g,h,i,j,k , Aruna Devi l, Anirudh Atmakuru m, Ru-San Tan n
a
School of Mathematics, Physics and Computing, University of Southern Queensland, Springfield Central, QLD 4300, Australia
b
School of Mathematics, Physics and Computing, University of Southern Queensland, Toowoomba QLD 4350, Australia
c
Cogninet AI, Sydney, NSW 2010, Australia
d
Faculty of Engineering and Information Technology, University of Technology Sydney, Sydney, NSW 2007, Australia
e
School of Computing and Information Science, Anglia Ruskin University Cambridge Campus, UK
f
Australian International Institute of Higher Education, Sydney, NSW 2000, Australia
g
School of Science Technology, University of New England, Australia
h
School of Biosciences, Taylor’s University, Malaysia
i
School of Computing, SRM Institute of Science and Technology, India
j
School of Science and Technology, Kumamoto University, Japan
k
Sydney School of Education and Social Work, University of Sydney, Australia
l
School of Education and Tertiary Access, University of the Sunshine Coast, Petrie, QLD 4502, Australia
m
University of Massachusetts, Amherst, USA
n
National Heart Centre Singapore, Singapore, Duke-NUS Medical School, Singapore

A R T I C L E I N F O A B S T R A C T

Keywords: Background and Objectives: Mild Cognitive Impairment (MCI) and Alzheimer’s Disease (AD) are progressive
Artificial intelligence neurological disorders that significantly impair the cognitive functions, memory, and daily activities. They affect
Neurological diseases millions of individuals worldwide, posing a significant challenge for its diagnosis and management, leading to
EEG signals
detrimental impacts on patients’ quality of lives and increased burden on caregivers. Hence, early detection of
Deep learning
Alzheimer’s disease (AD)
MCI and AD is crucial for timely intervention and effective disease management.
Mild cognitive impairment (MCI) Methods: This study presents a comprehensive systematic review focusing on the applications of deep learning in
detecting MCI and AD using electroencephalogram (EEG) signals. Through a rigorous literature screening process
based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, the
research has investigated 74 different papers in detail to analyze the different approaches used to detect MCI and
AD neurological disorders.
Results: The findings of this study stand out as the first to deal with the classification of dual MCI and AD
(MCI+AD) using EEG signals. This unique approach has enabled us to highlight the state-of-the-art high-per­
forming models, specifically focusing on deep learning while examining their strengths and limitations in
detecting the MCI, AD, and the MCI+AD comorbidity situations.
Conclusion: The present study has not only identified the current limitations in deep learning area for MCI and AD
detection but also proposes specific future directions to address these neurological disorders by implement best
practice deep learning approaches. Our main goal is to offer insights as references for future research encour­
aging the development of deep learning techniques in early detection and diagnosis of MCI and AD neurological
disorders. By recommending the most effective deep learning tools, we have also provided a benchmark for
future research, with clear implications for the practical use of these techniques in healthcare.

* Corresponding authors.
E-mail addresses: [email protected] (M. Acharya), [email protected] (R.C. Deo).

https://doi.org/10.1016/j.cmpb.2024.108506
Received 13 August 2024; Received in revised form 29 October 2024; Accepted 6 November 2024
Available online 12 November 2024
0169-2607/© 2024 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

1. Introduction have shown great potential to detect and diagnose MCI and AD based on
EEG signals [18]. Instead of relying entirely on human intervention,
Dementia is a group of diseases that manifest with a decline in which can be time-consuming and expensive, deep learning techniques
memory, reasoning, and the ability to perform daily activities. Globally, self-optimise by analysing large amounts of data to detect intricate
more than 55 million people suffer from dementia worldwide, with over patterns in EEG signals automatically. In other words, they can detect
60 % in low-and middle-income countries [1,2], and 10 million new subtle patterns in the data that are otherwise difficult for humans to
cases are diagnosed every year [3]. Mild cognitive impairment (MCI) is discern [19]. The variations in EEG signals and their frequency spectra
considered the intermediate stage between the cognitive changes seen in associated with diseases allow diagnostic discrimination [20]. As EEG
normal aging–e.g., occasional forgetfulness–and those associated with signals have small amplitudes (microvolts), subtle changes in the
dementia [4]. Unlike dementia, MCI patients often retain functional different channels are difficult to analyse, especially when the data is
capacity. The prevalence of MCI among adults aged 60 and above ranges huge [21].
from approximately 6.7 % to 25.2 % [5]. The risk decreases with age and Fig. 2 provides a visual summary of the landscape of systematic re­
higher levels of education, and it is more prevalent in men [6]. The most views in AD and MCI detection, highlighting the complementary focus of
common cause of dementia is Alzheimer’s disease (AD), a progressive different studies. Each review covers a specific aspect of neurological
neurodegenerative disease characterized by amyloid-beta peptide’s detection (EEG, MRI, progression from MCI to AD), contributing to a
accumulation in the brain [7]. The exact etiology of AD is unclear, holistic understanding of how deep learning and other techniques are
although multiple factors have been implicated, including advanced age advancing the detection of neurodegenerative diseases.
and gene [8]. Genetic variations in the APOE ε4 allele are associated Hence, computer-aided diagnosis of MCI and AD is needed in the
with AD and MCI [9]–some MCI patients are pre-clinical AD who presence of voluminous data. Indeed, artificial intelligence (AI) tech­
eventually progress to clinical AD with severe functional impair­ niques, especially deep learning, have shown great potential to detect
ment–but the presence of a culprit gene may not always result in and diagnose MCI and AD based on EEG signals [22]. Instead of relying
cognitive decline [10]. Finally, several chronic diseases and lifestyle entirely on human intervention, which could be time-consuming and
factors have been linked to cognitive impairment, e.g., diabetes, hy­ expensive, deep learning techniques self-optimize by analyzing large
pertension, hypercholesterolemia, obesity, depression, smoking, lack of amounts of data to identify intricate patterns in EEG signals automati­
physical exercise, and low education levels [11]. cally. In other words, they can detect subtle patterns in the data that are
There is no definitive test for Alzheimer’s disease (AD). The clinical otherwise difficult for humans to discern [23].
diagnosis requires comprehensive cognitive and neurological evalua­ In this review paper, we aim to perform a detailed systematic review
tions, neuroimaging such as magnetic resonance imaging (MRI) and of deep learning methods for diagnosing AD and MCI neurological dis­
positron emission tomography (PET), and electroencephalogram (EEG) orders using EEG signals. The remainder of the paper is organized as
[12]. In imaging studies, the hippocampus (the memory center) shows a follows: Section 2 details the information search methodology related to
gradual decrease in size with age, mild cognitive impairment (MCI), and AD and MCI, Section 3 presents the results and synthesis of the findings,
AD (Fig. 1). Imaging tests are expensive, time-consuming, and require Section 4 discusses the results in further detail, and Section 5 presents
expert interpretation [13]. the conclusions and recommendations for future research directions in
Fig. 1 demonstrates the progression of brain degeneration from adopting deep learning for AD and MCI neurological disorder detection.
normal aging to severe AD, highlighting the specific structural changes
that occur in key brain areas responsible for memory and cognitive 2. Material and methods
functions. The Fig. serves as a visual aid to understand how Alzheimer’s
disease manifests and worsens over time. EEG, which can continuously 2.1. Related reviews
map the brain’s surface electrical potentials via multichannel scalp
electrodes, provides spatial and highly resolved temporal information Our literature search demonstrated that there currently appears to be
about the functional activities of various brain regions and has been used a lack of systematic reviews focusing on deep-learning methods for EEG-
to detect diverse neurocognitive disorders, including Alzheimer’s Dis­ based AD and MCI diagnosis. Therefore, using different combinations of
ease (AD) [14,15]. EEG signals are particularly useful in such applica­ the search words “EEG”, “deep learning”, “Alzheimer OR AD OR MCI OR
tions as they capture both linear and non-linear dynamic information Mild Cognitive Impairment” and “Review” on PubMed and Google
[16]. Disease-associated variations in EEG signals and their frequency Scholar, we found 14 review studies (Table A. 1) that were most relevant
spectra allow for diagnostic discrimination [17]. However, because EEG for AD and MCI neurological disorder detection. The search shows that
signals have small amplitudes (measured in microvolts), subtle changes several studies [24,25,26] have reviewed different techniques to iden­
in different channels are difficult to analyse, especially when the data set tify the various neuropsychiatric disorders, including MCI, AD, Parkin­
is large. Hence, there is a need for computer-aided diagnosis of Mild son’s diseases, bipolar disorder, depression, etc, but none of those
Cognitive Impairment (MCI) and AD in the presence of voluminous data. studies have specifically focused on MCI or AD detection techniques
Indeed, artificial intelligence (AI) techniques, especially deep learning, (like deep learning), and utilising the particular data types (like EEG)

Fig. 1. Brain structure of the normal individuals compared to the MCI and AD patients.

2
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

Fig. 2. Comparison of our review paper with other review papers developed for automated detection of AD and MCI.

considered in this research. Some papers reviewed deep learning, EEG 3. Results
signals, and MCI or AD in general, but they lacked a specific focus on
EEG-based MCI/AD detection using AI techniques. [27] and [28] have Among the 74 papers published in the last decade, 45 papers (61 %)
also studied the application of machine learning (ML) to detect AD using have been published from 2022 (Fig. 4), which mirrors the secular
EEG images. Still, these studies did not include the high-performing deep development of state-of-the-art deep learning models in the field as well
learning models that are commonly used today. as computing power, which significantly improved the accuracy and
In Fig. 2, we depict how our carefully selected papers differed from efficiency of analyzing EEG signals for detecting MCI and AD. For the
five of the most recent related review studies in this field. analysis of results, we stratified the 74 papers into three groups based on
According to Fig. 3, there are three phases: Identification, Screening the condition/s being classified: MCI (11 papers), AD (38 papers), or
and Inclusion. Initially, 440 papers were collected, with 359 remaining both (25 papers). The first two groups classify MCI patients versus
after removing duplicates. In the Screening phase, 244 papers were healthy controls and AD patients versus healthy controls, respectively.
excluded for irrelevance, leaving 115 for assessment. Finally, 41 more The third group included papers that either classify MCI versus AD
papers were excluded due to a lack of machine learning or deep learning versus healthy controls or MCI versus AD patients. Among the last group
results, resulting in 74 papers being included in the review. The process are papers that studied the progression of MCI into AD.
ensures a rigorous selection of relevant, high-quality studies. Fig. 4 highlights a significant increase in research starting in 2018,
with peaks in 2022 and 2023 due to advances in deep learning models
2.2. Literature search and computational power. 2014-2017 saw low activity, likely reflecting
the initial stages of applying AI to EEG analysis. A drop in 2020 is
We conducted a comprehensive literature search in PubMed, Scopus, attributed to the pandemic, but there was a resurgence in 2021, aligning
Web of Science, and IEEE repositories for articles published between 1st with improvements in machine learning technologies. The text notes a
January 2014 and 30th June 2024, following the PRISMA guidelines. decline in publications in 2024, potentially indicating stabilization in
Using the Boolean string: (Alzheimer OR AD OR MCI OR mild cognitive research. Fig. 4 shows that 61 % of papers were published from 2022
impairment) AND (Deep Learning OR Transfer Learning OR Natural onwards, emphasizing the field’s rapid growth, driven by technological
Language Processing) AND (EEG OR Electroencephalogram), the initial advancements and the need for early detection of neurodegenerative
search yielded 440 results across the four databases. After removing 81 diseases.
duplicates and excluding 244 irrelevant works, review studies and non-
journal publications, 115 articles remained. Of these, 41 papers did not
report any performance results of the deep learning models, leaving 74 3.1. MCI detection
papers for analysis and review (Fig. 3: Article search strategy based on
the PRISMA guidelines.). In this section, we report the key papers that used deep learning
methods for EEG-based MCI diagnosis (Table A. 2). Early detection of
MCI is crucial so that intervention can be introduced to retard its

3
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

Fig. 3. Article search strategy based on the PRISMA guidelines.

Fig. 4. Distribution of articles by the publication year.

progression to AD, which is associated with severe cognitive decline, independent component analysis (ICA), and the grand average method
loss of independence, and higher healthcare costs. Modifications in on data acquired from subjects who had performed a P300 speller task.
brain connectivity density can be detected on the EEG. Compared with Their t-RIDE algorithm demonstrated high efficiency for MCI diagnosis
MCI, AD patients exhibit increased amounts of permutation Jaccard versus controls, which opens up the possibility of AI-based techniques
distances (PJD) and reduction in network density across all sub-bands for automatic MCI screening and, via mobile devices, remote monitoring
[29]. [30] investigated various techniques like tuned residue iteration of neurocognitive function.
decomposition (t-RIDE), residue iteration decomposition (RIDE), Later studies are broadly categorized into three groups: articles that

4
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

used scarce data, transfer learning models, and EEG data in a multi­ the EMCI index could effectively differentiate between MCI and HC
modal setting. To overcome the challenge of obtaining large open-access groups, with accuracy rates of 89.09 % and an F1 score of 89.44 % in the
EEG databases to train AI models, [31] has developed a model based on beta frequency band.
iterative amplitude adjusted Fourier transform (IAAFT) and bidirec­
tional long short-term memory (BiLSTM) that could distinguish MCI 3.2. AD detection
patients versus healthy controls using small amounts of EEG signal data.
Used along with data augmentation, IAAFT generates surrogate EEG In this section, we report the papers that used deep learning methods
data with similar characteristics to the original data, thereby decreasing for EEG-based AD diagnosis (Table A. 3). Deep learning techniques were
the chances of overfitting and improving generalizability. used for both feature extraction [17] and classification [40,41,42,43].
By additionally deploying BiLSTM to capture temporal dependencies [44] investigated various biomarkers, including EEG signals, priori­
in EEG signals, their model attained 97.20 % classification accuracy on a tizing non-invasiveness, cost, and portability. Applying multiscale
small dataset of 10 MCI patients and 10 healthy controls. Other studies analysis and embedding space theory to EEG-based brain functional
relied on pre-processing steps to mitigate data scarcity, including networks, they were able to classify AD patients versus healthy controls
denoising using techniques like stationary wavelet transformation, with 98 % accuracy using traditional classifiers. [45] developed a deep
segmentation, and down-sampling of raw data to expand the sample size learning model that could distinguish AD from healthy aging adults
and achieve similar classification performance. The key lies in per­ based on signal inputs acquired using a two-lead ambulatory EEG system
forming clever transformations to represent the available data in a more that recorded sleep-related data. Compared with healthy controls, AD
interpretable form. In [32] and [33], stationary wavelet and discrete patients spent less time in slow-wave sleep; other sleep stages were not
wavelet transformations, respectively, were used to decompose EEG significantly different between the two groups.
signals into sub-bands, from which features were extracted to perform Among the deep learning architectures, convolutional neural net­
MCI versus normal classifications, which attained accuracies exceeding works (CNNs), transfer learning models, and generative adversarial
95 %. These studies demonstrate that by carefully selecting EEG chan­ networks (GANs) have gained popularity for AD diagnostic applications.
nels and features, it is possible to achieve high accuracy in detecting CNNs are adept at image analysis and can automatically learn intricate
MCI, even in the presence of limited data. patterns from EEG signals. Moreover, they can extract features and
[34] used ResNet-18 to distinguish MCI patients versus healthy produce highly accurate results when combined with advanced graph-
controls based on the frequency and spatial properties of EEG data. They based networks [46,47,48,49]. [50] used CNNs to extract spatiotem­
observed that the frontal, left temporal, and parietal areas of the brain poral features from multi-channel time series EEG signal data efficiently.
were most affected in MCI patients, which differed from healthy controls However, they formatted the EEG signals collected from multiple brain
in the θ and low α bands on the EEG signal. From studying how MCI regions into a 2D array. This allowed them to analyze the in­
patients differed from mild AD patients, they observed that MCI patients terrelationships between these areas like in an image. Their model
had a larger affected right temporal area. Incorporating these findings attained 100 % classification accuracy for AD detection While such
into the model, they attained a high 98.33 % accuracy for the classifi­ complex models perform well, the results generated are unclear. In
cation of MCI patients versus healthy controls. [35] reproduced similar developing their CNN-based AD detection model, [48] incorporated
results on a larger, publicly available database, which not only lent explainability and interpretability by using saliency maps to visually
support to the results of [34], but also suggested a disrupted brain highlight components of the EEG signals (e.g., which frequency bands
network organization in MCI patients characterized by higher local ef­ and modulations) that contributed the most to model classification. This
ficiency in the beta band on the EEG signal. combination of CNNs and saliency maps not only improved diagnostic
Integrating information from multimodal sources in individual pa­ accuracy but also provided a clearer understanding of the model’s
tients can potentially provide a more complex and comprehensive cap­ decision-making process. In a recent study, [51] used CNNs to classify
ture of the disease for more a accurate diagnosis. However, it can come AD patients into two groups based on severity. Their CNN model used
at the cost of increased difficulty in acquiring the different data. Several multiple layers, including three convolutional and two fully connected
studies used such an approach to detect MCI, with EEG signals being part layers, to process 2D matrices extracted from multi-channel EEG signals,
of the multimodal datasets. [36] They used eye movement data, infor­ as in [50]. From a database of more than 650 patients from five different
mation from neuropsychological assessments, and EEG signals. Their hospitals, they classified AD patients into moderate and advanced with
model extracted 40 features using logistic regression, among which five over 97 % accuracy.
features were significantly related to MCI, yielding 81.51 % classifica­ Pre-trained on ImageNet data, transfer learning AD detection models
tion accuracy. [37] studied mindfulness impact in MCI using do not need to be trained from scratch and only require fine-tuning to
sleep-related information. In a double-blind randomized controlled trial, the specific EEG dataset. [52] employed AlexNet on resting-state EEG
they assigned 75 patients with MCI and insomnia into two groups: the signals to classify AD versus MCI versus MC healthy aging, attaining
mindfulness group and the health education control group. over 98 % three-class classification accuracy. This study showcased the
Compared to the control group, the mindfulness group showed sig­ potential of transfer learning for AD detection, which paved the way for
nificant improvements in sleep quality along with low levels of anxiety the development of more advanced transfer learning models. [53] used
and stress. Analysis of EEG recordings of both these groups revealed ResNet-50 to detect AD using EEG signals. In their model, raw EEG data
changes in brain activity, indicating relaxation and alterations in fre­ were first pre-processed using principal component analysis to remove
quency bands associated with attention during mindfulness practice. noise and then decomposed using wavelet transforms to extract statis­
This suggests that mindfulness can enhance sleep and cognitive abilities tical features. These features were converted into 2D plots, which then
in MCI patients, providing a low-cost, scalable intervention suitable for served as input to ResNet-50. To obviate the need for fine-tuning of
implementation. [38] studied robot-based training for improving pre-trained networks, some researchers proposed novel transfer learning
working memory and cognitive function in older patients. Their nu­ architectures trained specifically for AD detection to address the specific
merical results indicated that the intervention led to an 8 % increase in needs of the application. [54] proposed EEGAlzheimer’s Net, a transfer
cognitive scores, as measured by standardized assessments, compared to learning-based attention LSTM to handle non-linearities in EEG signals.
the control group. [39] developed an EEG-based MCI to assess cognitive The architecture combined spatial and temporal feature extraction using
workloads in MCI patients. They recorded EEG data from 124 brain CNNS and recurrent neural networks (RNNs), respectively, with a
areas of participants as they performed different cognitive tasks on a transformer-based architecture to detect AD. Model accuracy was 99.85
robot simulator. EEG-based MCI was found to be sensitive to changes in %. To address the spontaneous and highly variable nature of EEG sig­
the subject’s mental workload. Their study’s results demonstrated that nals, [55] proposed Adazd-Net, which used adaptive flexible analytic

5
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

wavelet transforms to capture subtle changes in EEG signals associated dependencies in data.
with AD. They then incorporated Shapely explanations, Morris sensi­
tivity analysis, and local interpretable model-agnostic explanations 3.3. MCI+AD (MCI and SD) detection
(LIME) by depicting the contributions of each feature towards
decision-making and providing explainability and interpretability to the In this section, we describe the studies that used EEG recordings and
model. Model accuracy was 99.85 %. deep learning for MCI+AD detection (Table A. 4), which encompassed
Several researchers have developed GANs and encoder-based models three-class classification of healthy versus MCI versus AD, binary clas­
for EEG-based AD diagnosis (but less so for MCI diagnosis). To address sification of MCI versus, or progression of MCI to AD. Deep learning
limited and imbalanced EEG training data in their AD detection models, setups like artificial neural networks can help eliminate noise and
[56] and [57] used GANs to generate synthetic data points very similar identify invariant features within these signals [68]. In addition, deep
to real data, which facilitated model training and improved model ac­ learning extracts and selects features automatically, which can either be
curacy. To improve the quality of synthetic data and produce explain­ classified by the deep model or be fed to standard ML classifiers. In their
able predictions, [56] used Wasserstein GAN to minimize the deep model, [68] used traditional ML algorithms like naïve Bayes for
earth-mover distance between real and synthetic data distributions. classification, attaining 98.25 % accuracy for three-class classification.
[57] combined GANs with a Marine Predator Algorithm to optimize Other researchers have also adopted this approach [69,70,71,72]. [69]
feature extraction and improve accuracy. Unlike GANs, which synthe­ used fast Fourier and continuous wavelet transforms to identify specific
size data, encoder-based models extract features and are widely com­ bands in the EEG signals that capture the most important features that
bined with CNNs, LSTMs, and transformers for AD detection. [58] discriminate AD versus MCI versus healthy controls.
proposed DICENet, a convolutional-transformer-based encoder archi­ Inputting deep learning-extracted features to a k-nearest neighbors
tecture, for AD diagnosis. DICENet comprises two parallel convolutional (KNN) classifier, which is simple yet able to handle non-linear data
blocks that perform dimensionality reduction, yielding outputs that are effectively, they attained 99 % model classification accuracy. [72] also
fed to a transformer-based model to make the final prediction. [59] used used KNNs for classification, but performed the feature selection
spatiotemporal autoencoders with CNNs and LSTMs to analyze brain differently. They introduced a technique called iterative filtering
dynamics as assessed by EEG in AD patients, attaining over 96 % AD decomposition to decompose EEG signals into intrinsic mode functions.
classification accuracy. They observed that their model was robust From here, four crucial features–power spectral density, Tsallis entropy,
against different instabilities of EEG signals and that the brain state variance, and fractal dimension–are extracted. Including other cognitive
trajectories of AD patients manifested as ring-manifolds, distinguishing tests to enrich these features, they attained 92 % accuracy. The impor­
them from controls. tance of efficient feature selection increases as the number of classes in
Like in the case of MCI detection, several recent studies [60,61,62] the classification task increases. Apart from neural networks and Fourier
have explored the use of multimodal data, information fusion, and transformations, ANOVA and regressions have also been used for feature
model fusion for EEG-based AD detection. These approaches typically selection in such cases [70]. In their model, the EEG features obtained
have high robustness and generalizability and produce better results using ANOVA and Ridge regression reflected phase, spectral, and tem­
than models that use data from one source alone. [63] analyzed EEG poral characteristics during rest and memory-encoding states. With such
data in combination with functional near-infrared spectroscopy (fNIRS), comprehensive feature extraction, they were able to successfully classify
which observed significantly better results in AD versus healthy classi­ the participants into four groups–AD, amnestic MCI, non-amnestic MCI,
fication performance than models trained on either EEG or fNIRS data and subjective cognitive decline–with 93.1 % accuracy.
alone. [64] studied the impacts of AD on spontaneous brain activity by While the works above involved deep learning for feature extraction/
using a thermodynamics-based framework to map the asymmetry of selection only, recent works have also applied deep learning for the
brain dynamics with time using a multimodal dataset of functional classification of MCI and AD from EEG signal data [73,74,75,76,77].
magnetic resonance imaging (fMRI) and EEG signals collected from AD [78] built an attention-based technique to distinguish MCI from sub­
patients and healthy controls. They found that AD correlated with a jective cognitive decline using resting state EEG signals. Owing to its
breakdown of temporal irreversibility at global, local, and network multi-head attention, the transformer architecture is particularly effi­
levels, and across multiple oscillatory frequency bands. The limbic, cient at handling temporal dependencies. By pre-processing EEG signals
frontoparietal, default mode and salience networks were particularly and extracting relevant frequency bands such as delta, theta, alpha, and
affected, while temporal irreversibility was linked to cognitive decline in beta, the model achieved an area-under-curve (AUC) of 0.807, demon­
AD and gray matter volume in healthy controls. [65] combined multi­ strating the potential of deep learning models for MCI detection. Deep
modal data with ensemble deep learning models. From EEG signals and learning and transfer learning models often outperform ML classifiers.
fNIRS collected concurrently during cognitive tasks, features derived [75] built a deep neural network model called CEEDNet that consisted of
from both modalities were optimized using a Pearson correlation different state-of-the-art models like VGG, ResNet, and vision trans­
coefficient-based feature selection strategy. Using a hybrid system that formers to detect MCI versus AD patients using spatial and temporal EEG
combined deep learning methods and decision analysis techniques, they signals and attained an AUC score of 0.9. While these models are
accurately classified subjects into healthy, MCI, and two levels of AD extremely powerful, data augmentation and extensive pre-processing
severity, showcasing the potential of integrated data to refine medical are often required to prevent this. [79] investigated the importance of
assessments. [66] built an ensemble model combining multiple data augmentation and its impact on transfer learning approaches. By
2D-CNNs to capture intricate patterns in EEG images. Each CNN model performing data augmentation on their data, they were able to improve
serves as an individual classifier, and the predictions from all CNNs are the classification accuracy of transfer learning models like ResNet by
aggregated to form the final decision. In so doing, the model mitigates over 5 %.
inherent model biases, attaining 97.9 % accuracy, which outperformed Among diverse deep learning architectures, CNNs are the most
the individual models. common for medical image analysis, often attaining classification ac­
[67] used the state-of-the-art vision transformers to screen for AD curacies exceeding 90 % [80,81,82]. [77] used CNNs to distinguish AD
patients. EEG images are first converted into scalograms and fed to the from healthy cognitive aging using EEG data and attained 92.5 % model
vision transformer model. These models treat each segment of the EEG accuracy. Of note, their work exploited CNNs in two ways. First, the
scalogram as part of a larger picture, assessing not only local features but CNN model allowed them to bypass the complexities of traditional
also how these features relate to others across the entire image, allowing feature engineering, directly identifying patterns in EEG signals associ­
for comprehensive extraction of spatial and temporal patterns. Further, ated with different dementia subtypes, such as AD, dementia with Lewy
the self-attention block allows the capture of complex long-range bodies, and idiopathic normal-pressure hydrocephalus. Second, the

6
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

model’s temporal resolution and ability to learn from short, overlapping Fig. 5 presents the scalable workflow for diagnosing cognitive im­
EEG segments enable it to capture transient neural dynamics disrupted pairments. By combining EEG data with deep learning and cloud tech­
in dementia. nology, this system enables automated and accessible detection of
Recent papers have also taken advantage of CNNs by using them in normal aging, mild cognitive impairment, and Alzheimer’s disease,
combination with other models. [83] proposed a hybrid model that used potentially leading to earlier and more accurate diagnoses.
CNN and bi-directional gated recurrent units to detect MCI. In this work, In the following sections, we will describe some of the best models
CNNs are cleverly used to extract both spatial and temporal data asso­ across three key areas: MCI detection, AD detection, and MCI + AD
ciated with EEG recordings. Specifically, the model uses small and large detection. Each domain addresses different aspects of neurodegenera­
temporal CNNs to independently capture temporal aspects of the EEG tive disease identification, providing unique insights into how deep
data, which are then combined. A spatial CNN further processes these learning can be applied to EEG data. We will also discuss two critical
combined features to exploit spatial relationships between different areas that have gained prominence recently: multimodality and
brain regions. Such a model is highly robust and [83] was able to detect explainability, both of which have substantially impacted the develop­
MCI patients with 99 % accuracy despite using images that had not been ment of more reliable and transparent AI models.
pre-processed. [84] and [85] used graph convolutional networks (GCNs) Multimodality refers to integrating multiple data sources, such as
to construct brain functional networks from EEG data for MCI and AD EEG, MRI, and PET scans, to enhance the diagnostic accuracy of deep
detection. GCNs effectively capture topological structures and neural learning models. While EEG provides valuable temporal information
interactions within EEG-derived functional connectivity, enabling the about brain function, combining it with structural data from MRI or PET
extraction of significant features and patterns associated with different scans offers a more comprehensive view of the brain, leading to
stages of dementia. Here, the use of graph theory metrics enhances improved detection of MCI and AD. This approach has become
diagnostic precision, highlighting the potential of GCNs in clinical ap­ increasingly popular as it helps to overcome the limitations of single-
plications for early and accurate MCI, AD, and dementia diagnosis. modality analysis, providing a richer understanding of
The performance of deep models can potentially improve with the neurodegeneration.
incorporation of multimodal data. Notably, deep architectures can Explainability, on the other hand, has emerged as a crucial factor in
process large amounts of data and extract useful features from all building trust with clinicians and patients. Traditional deep learning
different sources. As a result, building functional connectivity frame­ models are often considered "black boxes" that deliver predictions
works, using information fusion and developing diagnostic tools to without offering insight into how those predictions were made. Recent
analyze EEG signals automatically, have become popular [86,87,88]. advances in explainable AI (XAI) techniques, such as SHAP, LIME, and
[89] explored the use of complex tensor factorization with the PAR­ GradCAM, now allow researchers to interpret the decision-making
AFAC2 model for estimating brain connectivity from EEG data. Their process of these models. By highlighting which EEG features or brain
EEG model was built to effectively decompose EEG data into meaningful regions contribute most to a prediction, explainable models can provide
scalp components that are described by spatial, spectral, and complex clinicians with clearer insights into the workings of the model, making
trial profiles. They derived a new connectivity metric based on the them more likely to be adopted in clinical settings.
complex trial profiles of the extracted components and showed that In the following sections, we will explore the top-performing models
PARAFAC2 outperformed other traditional tensor analysis methods like in each of the three domains and how these models incorporate multi­
PARAFAC and MVAR-ICA. modal data and explainability techniques to improve the detection of
While dealing with feature fusion and multimodal data, one needs to MCI and AD.
be wary of incomplete data as it is extremely difficult to obtain complete
data from multiple sources for the same subject. [76] proposed a method 4.1. Key findings and discussion of the best models
to handle incomplete multimodal data using a disease-image-specific
deep learning framework that integrates image synthesis and disease From our review of 74 papers, we found that CNNs are widely
diagnosis into a unified process. It comprises a disease-image-specific adaptable and perform exceedingly well in different tasks ranging from
network (DSNet) for modelling disease-image specificity and a feature extraction to classification. They can also be used as ensembles
feature-consistency generative adversarial network (FGAN) for with other CNNs or deep learning architectures. Transfer learning
imputing missing images. While DSNet captures disease-relevant infor­ models like ResNet have performed the best in detecting MCI and AD
mation from whole-brain scans, the FGAN module is used for synthe­ and require minimal training time. Coming to MCI+AD detection,
sizing missing images while ensuring feature consistency. By training feature extraction is the most crucial step. This is because it might be
DSNet and FGAN together, they generated synthetic diagnosis-oriented difficult to distinguish between patterns of MCI and AD patients
images that achieved state-of-the-art performance in detecting AD as compared to MCI/AD patients versus healthy controls. In all three cases,
well as the progression of MCI to AD. [90] used a combination of EEG, multimodal models outperformed all others and demonstrated the po­
eye tracking, and behavioral data to provide a cost-effective and tential to use information from different sources to detect these diseases
noninvasive diagnostic alternative to traditional clinical methods, which at an early stage.
are often expensive and require specialized expertise. Using domain In the context of detecting MCI, four papers reported accuracy scores
adversarial neural networks and data augmentation, they attained ac­ (Fig. 6). Among these, the ResNet model by [34] performed the best.
curacies of 88.81 % and 100 % for MCI and AD diagnoses, respectively. EEG recordings were first processed to obtain low-order functional
connectivity (LOFC) scores, which helped quantify the interactions be­
4. Discussion tween different brain regions. The LOFC measurements from four fre­
quency bands were combined to create a multi-channel input for the
Deep learning has burgeoned due to recent exponential improve­ ResNet model. Because of ResNet’s ability to capture both local and
ments in computational power, data availability, and algorithmic in­ hierarchical features in the data, it identified patterns specific to MCI
novations. This has a direct impact on its applications in detecting MCI and analyze how different regions of the brain connected and commu­
and AD based on EEG recordings over the last couple of years (Figure). In nicated across various frequency bands. In so doing, the model attained
the following sections, we will describe some of the best models in each accuracy of up to 100 %.
of the three domains (MCI detection, AD detection, and MCI+AD In AD detection, we have observed that different authors quantified
detection) and highlight pressing aspects like multimodality and their results using different metrics. These results are summarized in
explainability that have gained the spotlight over the last couple of Figs 7- 9 in terms of accuracy, AUC, and F1 scores. Due to their well-
years. established ability to extract features and perform classification based

7
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

Fig. 5. Pictorial summary of MCI/AD/MCI+AD detection using deep learning and EEG signals.

train five 2D-CNNs independently on the same EEG dataset after pre-
processing and noise removal. Building an ensemble of five indepen­
dent CNNs allows each constituent model to independently identify and
learn from different features, potentially capturing a broader spectrum
of diagnostic signals. To make the final prediction, the outputs of each
individual CNN are aggregated to form a single, more accurate predic­
tion. This reduces the chances of overfitting and mitigates the risk of any
model introducing its bias into the prediction, thus guaranteeing
increased generalizability and robustness.
With respect to MCI+AD detection, several machine learning models
performed exceedingly well, achieving accuracies of over 90 % (Fig. 10).
Once again, different metrics were used to quantify model performances
(Figs. 11 and 12) . As discussed earlier, [69,90], and [83] are a few
examples, all of whom achieved over 99 % predictive accuracy. [86]
produced a similar work by developing a tool called a lacsogram to
Fig. 6. Accuracy scores for models used in detecting MCI. characterize MCI and different stages of AD. Their work relies on elab­
orate signal processing. EEG signals are first decomposed using discrete
on image data, many studies have used CNN-based models to detect AD. wavelet transforms into the delta, theta, alpha, beta, and gamma
Their work highlights the predictive ability of deep learning models in sub-bands since each of these might determine AD differently. Lapstral
an ensemble setup. The paper utilizes an ensemble approach where they and cepstral analyses are conducted on these signals and different

Fig. 7. Accuracy score for the models used in detecting AD.

8
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

Fig. 8. AUC scores for the models used in detecting AD.

Fig. 9. F1 scores for models used in detecting AD.

distance measures are used between these lapstral and cepstral co­ counterparts and that performing elaborate signal processing, data
efficients of different sub-bands. These distance measures then help augmentation, and feature selection can help build reliable, robust
statistical and ML models quantify the differences between the EEG models, in all the cases of MCI, AD, and MCI+AD detection. In partic­
patterns of healthy controls and patients with different stages of AD. ular, ResNet has been the best-performing model for MCI classification,
While these results highlight the high performance of deep learning while an ensemble of CNNs has been best suited for AD detection. In the
models, collectively these works suggest that multimodal and ensemble more complex case of MCI+AD detection that typically deals with a
models always stand out and perform significantly better than their three-class classification problem, feature extraction is key, and the

9
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

Fig. 10. Accuracy scores for the models used in detecting MCI+AD.

Fig. 12. F1 scores for models used in detecting MCI+AD.

Fig. 11. AUC scores for models used in detecting MCI+AD.

identifies all non-AD patients. DCssCDBM model, GAN + MPA, and DEL
model’s predictive ability depends on how well it distinguishes between
also exhibit very high specificity (>99 %), making them excellent for
MCI and AD features which are likely to be more similar as compared to
avoiding false positives. The ViT model shows the lowest specificity
that of healthy controls. Further, while accuracy and AUC scores have
(57.10 %), indicating it struggles with correctly identifying non-AD
been commonly used metrics to report model performance, sensitivity
patients, resulting in a high rate of false positives.
and specificity scores have also been used in a few works (Fig. 13).
MOCA achieves perfect sensitivity (100 %), correctly identifying all
Moving ahead, we urge the research community to adopt a single reli­
MCI+AD cases, making it the best-performing model for combined
able metric, such as the accuracy of F1 scores, to allow for model
detection. ResNet (93.33 %) and DSDL (91.05 %) also perform well, but
comparison and reproduction of results.
models like PCA + FBCSP show slightly lower sensitivity (87 %).
The Multi-Modal Classification Method achieves the highest sensi­
MOCA again leads in specificity (97.38 %), meaning it avoids false
tivity (100 %), meaning it correctly identifies all AD cases. STCGRU and
positives better than the other models. ResNet and DSDL also perform
KNN models also perform excellently, with sensitivity scores close to 99
well with specificity scores above 88 %, but PCA + FBCSP shows the
%. Lacsogram, MOCA, and other models also show very high sensitivity
lowest specificity (80 %), meaning it is less effective in identifying non-
scores (>97 %), suggesting robust performance in identifying AD pa­
MCI/AD patients compared to the top models.
tients. The Feature Fusion Model has the lowest sensitivity (86.08 %),
The findings of the reviewed models according to Figs 13 -- 16 are:
indicating it misses more AD cases than other models.
Adazd-Net shows perfect specificity (100 %), meaning it successfully

10
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

Fig. 13. Sensitivity scores for models used in detecting AD.

Fig. 14. Specificity scores for models used in detecting AD.

- Models like MOCA, Multi-Modal Classification Method, Adazd-Net, 4.2. Benefits and challenges of multi-modality
and STCGRU show consistently high sensitivity and specificity,
making them the best choices for both AD and MCI+AD detection. With the deep learning models processing large amounts of data
- While models like the Multi-Modal Classification Method and MOCA easily and identifying the crucial features from all of them, multimodal
excel in sensitivity, others like Adazd-Net and GAN + MPA stand out data have gained popularity recently. Since these datasets involve
for their high specificity. integrating information from different sources about the same person,
- Lower-performing models such as the Feature Fusion Model and ViT they provide a holistic view of the person’s condition, thus enabling
model may require further refinement to enhance both sensitivity more reliable predictions. Further, since information is available at
and specificity, ensuring they can reliably detect AD and MCI cases different levels about the same individual, extracting different features
without producing false positives. from each of these would not only enable early detection of the disease,
but also open possibilities for personalized treatments. One such

11
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

Fig. 15. Sensitivity scores for models used in detecting MCI+AD.

used eye-movement data along with EEG recordings to detect MCI and
AD. Tracking the movement of eyes can give insights into the in­
dividual’s cognitive load and response accuracy, which may be dimin­
ished in the presence of MCI/AD. For example, fixation and saccade
tasks can help measure how quickly someone can redirect their gaze.
Difficulty in performing such a task can be a sign of execution
dysfunction common in MCI and AD patients. As a result, both [36] and
[90] noted that tracking eye movement in the presence of visual stim­
ulus can help ameliorate the performance of models that use EEG re­
cordings alone.
While multimodal data allows for the training of more robust and
generalizable models that are not overfitted to any setting, they come
with a separate set of challenges. Integrating data from multiple mo­
dalities into one deep learning model is a complex task and requires
technical expertise. Next, the complexity of multimodal models also
requires more computational and storage resources to process all the
information carefully. Moreover, obtaining information about the same
individual from multiple data sources is extremely difficult, tedious, and
time-consuming. In such cases, incomplete data should be dealt with
carefully [76]. Finally, interpreting the results obtained from such
models is extremely difficult because it is very challenging to identify
Fig. 16. Specificity scores for models used in detecting MCI+AD. which features from which data source were instrumental in making a
prediction.
example is the study by [62]. They used EEG recordings in combination
with PET scans, sleep measurement techniques, and cognitive testing to
4.3. Explainable AI and interpretability of deep learning models
study MCI and AD patients.
Among the 74 studies reviewed in this work, nine of them used
In the context of medical image analysis, deep learning models
multimodal data, among which five were published in 2022 or later.
appear to lack interpretability and explainability aspects because their
Among these studies, we found that EEG data was often used in com­
decision-making process involves complex, nonlinear computations
bination with other neuroimaging data like MRI or eye movement data.
across multiple layers, making it difficult to trace how different inputs
One such example is that of [64], who used EEG in combination with
and features affect prediction. Several works have highlighted the
fMRI data to study the temporal irreversibility of brain dynamics in AD.
importance of building explainable and interpretable deep learning
In this process, while fMRI provided spatial details about brain activity,
models in medical image analysis [91,92]. It is extremely crucial to build
they exploited EEG signals to get insights into temporal changes. By
such transparent models to gain medical practitioners’ trust and ensure
studying the irreversibility of time series signals across the two modal­
the application of deep learning in practice [93]. By highlighting the
ities, they found that AD patients showed significantly lower levels of
most discriminative features and ranking them in terms of their
complexity and temporal asymmetry in brain activity, which could be a
importance toward the final decision-making, such efforts add a layer of
crucial feature in differentiating them from healthy individuals. While
interpretability and support clinicians in decision-making [56]. This
[62] and [64] used different neuroimaging recordings, [36] and [90]
would not only provide enhanced diagnostic accuracy but improve

12
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

scalability and facilitate personalized treatment.

In the context of explainability and interpretability, visualization is
key. Visualization helps researchers understand how a model has made a
particular prediction and what features were given importance during
this decision-making process. For example, [55] used Shapley explana­
tions, Morris sensitivity analysis, and LIME to provide interpretability.
While Morris sensitivity analysis allows one to identify the input vari­
ables that are most influential, LIME and Shapley scores help understand
which features contribute the most to decision-making. They do so by
quantifying the contribution of each feature to the prediction of a spe­
cific instance. The visual attention mechanism adopted by [48] works
similarly. By incorporating visual attention within neural networks, one
can highlight those regions of the input EEG data that are most influ­
ential in making a prediction. Highlighting such features as specific time
points or electrode readings provides relevant understanding for clini­
cians using them. [94] show a different approach to achieve interpret­
ability. They represent EEG signals as a graph, where each node refers to
an EEG electrode and each edge denotes functional connections and
interactions of the brain. Following this, they build a gated graph con­
volutional model that allows for dynamic weighting of the edges and
nodes in the graph. Interpretability is achieved by studying these
weights since each edge corresponds to the brain’s functional connec­
tions and priority is given to those nodes that carry more predictive
power in predicting if a patient is healthy or affected by MCI or AD.
Several studies have attempted to make deep learning models more
explainable by identifying specific visual features or input components
that contribute most to the decision-making process in EEG-based MCI
Fig. 17. Distribution of private and public databased used.
and AD detection. These features are critical in improving the diagnostic
accuracy and usability of models in clinical practice.
The visual and EEG features that have been found to be most studies. Out of the 74 studies reviewed, 67 papers have explicitly cited
indicative of class discrimination between healthy controls, MCI pa­ the source of their data. Of these 67 studies, 49 used private datasets,
tients, and AD patients include: while 18 used publicly available datasets. Fig. 17 and Fig. 18 depict the
percentage of papers that used public and private datasets in each
- Specific electrode readings from temporal, parietal, prefrontal, and category.
hippocampal regions. Fig. 18 showcases the reliance on private datasets across AD and MCI
- Changes in functional brain connectivity observed through graph- detection studies. This trend points to a critical need for more publicly
based models. accessible datasets to support transparency, reproducibility, and inno­
- Important time points in EEG recordings, often in the alpha and beta vation in this domain. The scarcity of public datasets limits the research
bands. community’s ability to build on existing studies, highlighting a gap that
- Frequency bands such as theta and delta, which have shown to future research should aim to address. It is crucial to highlight that
correlate with cognitive decline. compared to studies that have used publicly available datasets, almost
- Spatial EEG topography, particularly in the occipital and frontal 2.5 times the number of studies have used private datasets. This in­
lobes, which helps distinguish AD patients. dicates the lack of large public datasets easily accessible to researchers

By identifying and highlighting these discriminative features using

explainable AI techniques like Shapley values, LIME, and Grad-CAM,
researchers have made deep learning models more transparent. This
interpretability is essential in clinical settings, where understanding
which specific features contribute to diagnoses allows clinicians to make
more informed decisions, ultimately enhancing the models’ adoption in
practice.
Among the 74 reviewed studies, considering that only four papers
[55,94,48,56] have worked on building explainable and interpretable
models, the working of all other papers remains a black box. We thus
urge future researchers to expand on such works to bridge the gap be­
tween using deep learning for MCI and AD detection in theory and
practice.
Fig. 17 highlights the need for more public dataset usage and
disclosure in the field of EEG-based MCI and AD detection. Coupled with
the call for more explainable models, this figure suggests a need for
greater openness and accessibility in future research efforts to foster
collaboration and innovation.

4.4. Databases used

In this section, we discuss the databases used by the reviewed Fig. 18. Distribution of databases based on the detection task.

13
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

in this domain. Analyzing this further, Fig. 19 sheds more light on this MCI, and AD. This multimodal approach facilitates comprehensive
limitation. Among the 67 studies that cited their data sources and sta­ research in studying potential biomarkers of both AD and MCI. It not
tistics, 21 works used datasets consisting of 50 or fewer participants, and only provides multimodal information, but also facilitates the study of
more than half the studies were based on datasets consisting of 100 or MCI, AD, and healthy controls simultaneously.
fewer participants. To build models that are robust, generalizable, and Table 1 provides a comprehensive overview of publicly available
capable of catering to different individuals in the real world, it is EEG datasets used for MCI and AD detection, which are critical for
extremely important to release publicly available datasets consisting of a advancing research in this field. These datasets offer a range of sample
large number of data points. sizes, conditions (MCI, AD, dementia), and multimodal features (e.g.,
Fig. 19 showcases the variability in sample sizes across studies in AD, MRI, PET), providing valuable resources for training and validating
MCI, and combined detection (AD + MCI) categories. While AD detec­ machine learning models aimed at early detection and progression
tion studies are more likely to use larger datasets, MCI detection studies monitoring of neurodegenerative diseases. The availability of these
often rely on smaller sample sizes. The figure underscores the need for datasets fosters collaboration, transparency, and reproducibility in the
larger, more diverse datasets, especially in MCI detection, to improve research community.
the generalizability and robustness of findings in this important area of Moreover, the deep learning models comparisons have been given in
neurodegenerative research. Table 2.
Table 1: Details of publicly available datasets for MCI/AD detection
using EEG recordings. Among these publicly available datasets, some of 4.5. Future research work
the largest ones include the Alzheimer’s Disease Neuroimaging Initiative
(ADNI) dataset, the Sina and Nour Hospital dataset [95], and the Florida To overcome the different limitations discussed earlier, we propose
State University dataset [96]. The Florida State University dataset re­ the following future research works:
cords EEG signals from healthy controls in two settings. The 96 partic­
ipants are divided into two categories of 46 individuals each. The first 1. Availability of a huge public database: It may be noted from this work
category consists of 24 AD patients and 24 healthy controls whose EEG that most of the research has been carried out using smaller private
signals are recorded with their eyes open. The second category consists or public databases. We propose to have more public databases for
of a similar split of participants, but the EEG signals are recorded with researchers to develop accurate, robust, and faster DL models. The
their eyes closed. Recent studies like [53] and [66] have used this public databases developed using data collected from various coun­
dataset to detect AD. The Sina and Nour Hospital dataset [95] has also tries and centers can aid in creating a robust model.
been widely used over the last couple of years [33,32,82,83]. It is a 2. Explainable AI and uncertainty quantification: AI models perform like
relatively smaller dataset consisting of EEG recordings of 27 subjects (16 black boxes by diagnosing input data. It does not explain the process
normal and 11 MCI participants) aged between 60 and 77. All EEG involved in obtaining output. Hence, explainable AI (XAI) can be
signals were recorded continuously using 19 electrodes in the morning employed to develop the confidence of clinicians and researchers.
with the participants resting comfortably in a quiet room with closed Such techniques can be employed in hospitals for the detection of
eyes. Because of the targeted age group of the participants, this dataset various mental disorders and treatment. Techniques such as LIME,
can be used to train models to distinguish between naturally aging in­ SHapley Additive exPlanations (SHAP), and Gradcam (Gradient-
dividuals and MCI patients. Although used in several research works, weighted Class Activation Mapping) have been developed to address
these databases are limited by the availability of EEG recordings and the model explainability to explain the working of the generated
cater only to the study of MCI or AD. On the other hand, the ADNI model [106].
dataset consists of a huge EEG dataset along with MRI and PET scans for 3. Uncertainty Quantification (UQ): Most developed AI models perform
studying AD and MCI. The EEG images in the dataset are used to assess well using small databases. Their performances vary when subjected
brain activity patterns that may differentiate between normal aging, to real-world scenarios in the presence of noises due to changes in the

Fig. 19. Studies stratified based on the size of the dataset.

14
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

Table 1 models are the structures fed as input to the DL models. DMTLM
Details of publicly available datasets for MCI/AD detection using EEG performs multiple trained tasks simultaneously. FL is a technology
recordings. that obtains information about the data for new AI models without
Database Name/Citation/ Studies Important characteristics touching it. DMLM improves each other’s performance among many
Link using the networks.
database 6. Data Standardization: During the data collection of EEG and other
ADNI dataset (https://adni. [76,35,64] ADNI primarily focuses on MRI physiological signals from various centers, due to the variation in the
loni.usc.edu/) and PET imaging, but it includes acquisition protocols and types of equipment, there is a possibility of
EEG data for multimodal MCI
changes in the magnitude of input data to the AI system. Such in­
and AD research.
Sina and Nour Hospital, [32,82,33, EEG dataset of 27 subjects - 16 consistencies introduce errors in the AI systems. To overcome these
Isfahan, Iran [95] (https:// 83] healthy controls and 11 MCI problems, data standardization needs to be done before feeding to
misp.mui.ac.ir/en/ee patients, all of whom are aged the AI system [117].
g-data-0) between 60 and 77 years. 7. Incorporation of Transfer Learning: Transfer learning can be explored
[97] [82] Consists of 109 subjects - 7 MCI
patients and 102 healthy
to address the issue of limited data availability in smaller databases.
controls This method allows leveraging pre-trained models developed on
Chung-Ang University Hospital [75] 1379 EEG recordings from 1155 larger, related datasets to improve model performance on smaller
EEG – CAUEEG Dataset [75] patients including normal (459), datasets. Transfer learning has shown promise in various fields,
MCI (417), Dementia (311)
including medical imaging, where models trained on larger datasets
classes.
The MCI and mild AD dataset – [79] MCI Dataset - consists of 22 MCI can be fine-tuned for specific tasks like MCI and AD detection using
[98,99] subjects and 38 healthy EEG data.
controls. Mild AD Dataset - 8. Personalized AI Models: Developing personalized AI models tailored
consists of 17 mild AD subjects to individual differences in brain activity can significantly enhance
and 24 healthy controls
the accuracy and reliability of MCI and AD detection. EEG signals
VSTMBT memory task EEG [89] EEG data collected based on the
Data [100] VSTMBT memory task from 23 exhibit high variability between individuals, so AI models that ac­
MCI patients and 24 healthy count for personalized baselines and patterns may yield better results
controls than generalized models.
OpenNeuro Dataset – [101] [102] EEG dataset of 36 AD patients,
9. Real-Time Detection and Monitoring: Future research can focus on
23 Dementia patients, and 29
healthy controls. developing AI systems capable of real-time EEG analysis for contin­
Scalp EEG dataset [50] Information from 19 channels uous monitoring of MCI and AD progression.
(https://osf.io/download/yh for 24 healthy individuals and
g9w/) 24 AD patients. 5. Conclusion
Hospital das Clínicas in São [52,103] EEG recordings from 31 mild AD
Paulo, Brazil [17] patients, 20 moderate AD
patients, and 35 healthy This review paper has analysed the various models employed for AD,
individuals. MCI, and (MCI+AD) categories from (2013-2024) using EEG signals. We
RFGHCPLA Dataset [104] [105] EEG recordings from 39 subjects have observed that the ensemble CNN model yields the highest for
(20 healthy controls and 19
automated detection of AD, ResNet is effective for detecting MCI, and
aMCI patients).
Florida State University dataset [53,66] EEG recordings taken with the efficient feature extraction using CNNs is extremely crucial for (MCI­
[96] eyes of individuals opened and +AD) detection. Further, we observed that multimodal datasets help
closed - 48 healthy controls and build robust, generalizable, and high-performing models in all three
48 AD patients. cases.
Matouˇs Cejnek dataset - [53] EEG recordings of 7 MCI
The limitation of this work is that most of the studies have used
(https://figshare.com/ patients, 59 AD patients, and
articles/ 102 healthy controls. smaller databases for the automated detection of classes based on EEG
dataset/dataset_zip/ signals. We need to use large databases from many countries and various
5450293/1) centers to overcome this limitation. Also, HRV or PPG signals can be
AD Classification dataset – [54] EEG data consisting of MCI
extracted using wearable devices to develop DL models and can be used
(https://github.com/ patients, AD patients, and
tsyoshihara/Alzheimer-s- healthy controls.
in home-based environments.
Classification-EEG/tree/ XAI and UQ must be employed to use the developed model in the
master/data) clinical environment. The deep learning model needs to be developed
using a huge, diverse data population belonging to various races and age
groups.
data or tuning parameters. In such cases, UQ can be employed to
quantify and mitigate the uncertainty in the data and model in the
Ethical approval
clinical environment [107,108,109].
4. Information Fusion: The performance obtained using EEG signals and
Not applicable.
DL models has been considered in this review. Physiological signals
such as electrocardiogram (ECG), heart rate variability (HRV),
Funding
photoplethysmography (PPG), or/and speech signals can also be
considered with brain images. Such data fusion methods can improve
The first author has received the University of Southern Queensland
the performance of the DL models with huge databases [110].
(UniSQ) Domestic PhD Research Scholarship (2023-2026) and the
5. Developing new DL models: New DL models, such as deep attention
Research and Training Program (RTP) Scholarship funded by the
mechanism models (Hafiz et al., [111]), graph convolutional neural
Australian Government, both of which are greatly acknowledged.
networks (GCNN) [112], deep multi-task learning models (DMTLM)
[113], federated learning (FL) models [114], and deep mutual
CRediT authorship contribution statement
learning models (DMLM) [115,116], can be explored to obtain ac­
curate performance with huge databases. The attention mechanism
Madhav Acharya: Writing – original draft, Visualization, Valida­
helps to focus on the important portion of the input. The graph
tion, Methodology, Investigation, Funding acquisition, Formal analysis,

15
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

Table 2
Comparative results of the deep learning models.
Model Architecture features Strengths Common applications Performance highlights

CNN (Convolutional Convolutional and pooling Strong feature extraction from spatial Feature extraction and Commonly used in AD detection, shows high
Neural Network) layers, fully connected layers data, effective for image-like EEG classification of EEG performance in ensemble learning (e.g., 99
inputs signals % sensitivity)
ResNet (Residual Residual connections, deep Handles very deep networks without MCI detection, EEG data Achieved up to 100 % accuracy in MCI
Neural Network) architecture vanishing gradient issues, captures both with multi-channel input detection
local and global features
LSTM (Long Short- Recurrent architecture with Captures temporal dependencies, Sequential EEG data, Paired with CNN for high specificity (96.33
Term Memory memory cells and gates retains information across sequences temporal analysis %) in AD detection
Network)
GCN (Graph Graph convolution layers, Models spatial relationships between EEG as brain High accuracy due to effective modeling of
Convolutional handles graph-structured data brain regions connectivity graphs, MCI brain connectivity
Network) detection
Autoencoders (Conv- Encoder-decoder structure, Learns compressed representations of Feature extraction and Achieved high sensitivity (92 %) in AD
AE) unsupervised feature learning EEG signals dimensionality reduction detection
Attention-based Attention layers, context Focuses on important parts of the input MCI and AD detection, Models like STCGRU achieve sensitivity
Models vectors data, enhances interpretability explainable AI close to 99 %
Federated Learning Distributed local models, Protects data privacy, enables training EEG data from Promising for privacy-sensitive EEG data,
(FL) central aggregation across multiple sources decentralized datasets reduces data sharing concerns
Multi-modal Models Parallel input streams, fusion Combines complementary information Combining EEG with Improved diagnostic accuracy through
layers for multiple data types (e. from multiple sources neuroimaging, AD modality integration (e.g., 99 % predictive
g., EEG + MRI) detection accuracy in combined MCI+AD detection)

Data curation, Conceptualization. Ravinesh C Deo: Writing – review & [10] E. Arnáiz, O. Almkvist, Neuropsychological features of mild cognitive impairment
and preclinical Alzheimer’s disease, Acta Neurol. Scand. 107 (2003) 34–41.
editing, Supervision, Resources, Project administration. Xiaohui Tao:
[11] E. Shatil, Does combined cognitive training and physical activity training enhance
Writing – review & editing, Supervision. Prabal Datta Barua: Writing – cognitive abilities more than either alone? A four-condition randomized
review & editing, Supervision, Resources, Project administration. Aruna controlled trial among healthy older adults, Front. Aging Neurosci. 5 (2013) 8.
Devi: Writing – review & editing, Supervision, Project administration. [12] Y.F. Tai, P. Piccini, Applications of positron emission tomography (PET) in
neurology, J. Neurol. Neurosurg. Psychiatry 75 (2004) 669–676.
Anirudh Atmakuru: Writing – review & editing, Resources, Method­ [13] F. Massa, M. Bauckneht, E. Peira, C. Lapucci, A. Picco, S. Capitanio, D. Arnaldi,
ology, Investigation. Ru-San Tan: Writing – review & editing, Super­ L. Roccatagliata, A. Chincarini, F. Nobili, Neuroimaging findings in mild
vision, Resources, Methodology. cognitive impairment, PET and SPECT in Neurol. (2021) 367–425.
[14] J. Dauwels, F. Vialatte, A. Cichocki, Diagnosis of Alzheimer’s disease from EEG
signals: where are we standing? Curr. Alzheimer. Res. 7 (2010) 487–505.
Declaration of competing interest [15] C.M. Michel, M.M. Murray, Towards the utilization of EEG as a brain imaging
tool, Neuroimage 61 (2012) 371–385.
[16] G. Rodriguez-Bermudez, P.J. Garcia-Laencina, Analysis of EEG signals using
The authors declare that they have no known competing financial nonlinear dynamics and chaos: a review, Appl. Math. Inf. Sci. 9 (2015) 2309.
interests or personal relationships that could have appeared to influence [17] R. Cassani, T.H. Falk, Alzheimer’s disease diagnosis and severity level detection
based on electroencephalography modulation spectral “patch” features, IEEE J.
the work reported in this paper.
Biomed. Health Inform. 24 (2019) 1982–1993.
[18] L. Lazli, M. Boukadoum, O.A. Mohamed, A survey on computer-aided diagnosis of
Supplementary materials brain disorders through MRI based on machine learning and data mining
methodologies with an emphasis on Alzheimer disease diagnosis and the
contribution of the multimodal fusion, Appl. Sci. 10 (2020) 1894.
Supplementary material associated with this article can be found, in [19] D.T. Nguyen, W. Li, P.O. Ogunbona, Human detection from images and videos: A
the online version, at doi:10.1016/j.cmpb.2024.108506. survey, Pattern. Recognit. 51 (2016) 148–175.
[20] S. Siuly, Y. Li, Y. Zhang, EEG signal analysis and classification, IEEE Trans. Neural
Syst. Rehabilit. Eng. 11 (2016) 141–144.
References [21] A. Puce, M.S. Hämäläinen, A review of issues related to data acquisition and
analysis in EEG/MEG studies, Brain Sci. 7 (2017) 58.
[1] M. Abd Al-Alim, R. Mubarak, N.M. Salem, I Sadek, A machine-learning approach [22] S.-A. Sadegh-Zadeh, E. Fakhri, M. Bahrami, E. Bagheri, R. Khamsehashari,
for stress detection using wearable sensors in free-living environments, Comput. M. Noroozian, A.M. Hajiyavand, An approach toward artificial intelligence
Biol. Med. 179 (2024) 108918. Alzheimer’s disease diagnosis using brain signals, diagnostics 13 (2023) 477.
[2] J. Hugo, M. Ganguli, Dementia and cognitive impairment: epidemiology, [23] P. Matthew, M. Anderson, Novel approaches to developing multimodal biometric
diagnosis, and treatment, Clin. Geriatr. Med. 30 (2014) 421–442. systems with autonomic liveness detection characteristics, in: Intelligent Systems
[3] A. Wimo, B. Winblad, H. Aguero-Torres, E. von Strauss, The magnitude of for Science and Information: Extended and Selected Results from the Science and
dementia occurrence in the world, Alzheimer Dis. Assoc. Disorders 17 (2003) Information Conference 2013, Springer, 2014, pp. 121–138.
63–67. [24] S. Chen, T.-L. Liu, Y. Jia, J. Li, Recent advances in bio-integrated electrochemical
[4] R.C. Petersen, Mild cognitive impairment, Continuum: lifelong Learn. Neurol. 22 sensors for neuroengineering, Fundam. Res. (2023).
(2016) 404–418. [25] F. Mattioli, V. Maglianella, S. D’Antonio, E. Trimarco, D. Caligiore, Non-invasive
[5] S. Jongsiriyanyong, P. Limpawattana, Mild cognitive impairment in clinical brain stimulation for patients and healthy subjects: Current challenges and future
practice: a review article, Am. J. Alzheimer’s Dis. Other Dementias® 33 (2018) perspectives, J. Neurol. Sci. (2023) 122825.
500–507. [26] M.J. Rivera, M.A. Teruel, A. Mate, J. Trujillo, Diagnosis and prognosis of mental
[6] A.M. McGrattan, Y. Zhu, C.D. Richardson, D. Mohan, Y.C. Soh, A. Sajjad, C. Van disorders by means of EEG and deep learning: a systematic mapping study, Artif.
Aller, S. Chen, S.-M. Paddick, M. Prina, Prevalence and risk of mild cognitive Intell. Rev. (2022) 1–43.
impairment in low and middle-income countries: a systematic review, [27] S. Yang, J.M.S. Bornot, K. Wong-Lin, G. Prasad, M/EEG-based bio-markers to
J. Alzheimer’s Dis. 79 (2021) 743–762. predict the MCI and Alzheimer’s disease: a review from the ML perspective, IEEE
[7] P. dos Santos, C. Leide, P.F. Ozela, M. de Fatima de Brito Brito, A.A. Pinheiro, E. Trans. Biomed. Eng. 66 (2019) 2924–2935.
C. Padilha, F.S. Braga, d.S. de Paula, H. Carlos, C.B.R dos Santos, Alzheimer’s [28] K.D. Tzimourta, V. Christou, A.T. Tzallas, N. Giannakeas, L.G. Astrakas,
disease: a review from the pathophysiology to diagnosis, new perspectives for P. Angelidis, D. Tsalikakis, M.G. Tsipouras, Machine learning algorithms and
pharmacological treatment, Curr. Med. Chem. 25 (2018) 3141–3159. statistical approaches for Alzheimer’s disease analysis based on resting-state EEG
[8] A.C. Naj, G.D. Schellenberg, A.s.D.G. Consortium, Genomic variants, genes, and recordings: A systematic review, Int. J. Neural Syst. 31 (2021) 2130002.
pathways of Alzheimer’s disease: An overview, Am. J. Med. Genet. Part B: [29] N. Mammone, C. Ieracitano, H. Adeli, A. Bramanti, F.C. Morabito, Permutation
Neuropsychiatr. Genet. 174 (2017) 5–26. Jaccard distance-based hierarchical clustering to estimate EEG network density
[9] S. Cervantes, L. Samaranch, J.M. Vidal-Taboada, I. Lamet, M.J. Bullido, A. Frank- modifications in MCI subjects, IEEE Trans. Neural Netw. Learn. Syst. 29 (2018)
García, F. Coria, A. Lleó, J. Clarimón, E. Lorenzo, Genetic variation in APOE 5122–5135.
cluster region and Alzheimer’s disease risk, Neurobiol. Aging 32 (2011) 2107,
e2107-2107. e2117.

16
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

[30] D. De Venuto, V.F. Annese, G. Mezzina, Remote neuro-cognitive impairment classification of Alzheimer’s disease using electroencephalogram, J. King Saud
sensing based on P300 spatio-temporal monitoring, IEEE Sens. J. 16 (2016) Univ. Comput. Inf. Sci. 35 (2023) 101848.
8348–8356. [58] A. Miltiadous, E. Gionanidis, K.D. Tzimourta, N. Giannakeas, A.T. Tzallas, DICE-
[31] X. Li, H. Zhou, R. Su, J. Kang, Y. Sun, Y. Yuan, Y. Han, X. Chen, P. Xie, Y. Wang, net: a novel convolution-transformer architecture for Alzheimer detection in EEG
A mild cognitive impairment diagnostic model based on IAAFT and BiLSTM, signals, IEEE Access (2023).
Biomed. Signal. Process. Control 80 (2023) 104349. [59] L. Wu, Q. Zhao, J. Liu, H. Yu, Efficient identification of Alzheimer’s brain
[32] A.M. Alvi, S. Siuly, H. Wang, K. Wang, F. Whittaker, A deep learning based dynamics with Spatial-Temporal Autoencoder: A deep learning approach for
framework for diagnosis of mild cognitive impairment, Knowl. Based. Syst. 248 diagnosing brain disorders, Biomed. Signal. Process. Control 86 (2023) 104917.
(2022) 108815. [60] N. Houmani, F. Vialatte, E. Gallego-Jutglà, G. Dreyfus, V.-H. Nguyen-Michel,
[33] M. Aljalal, S.A. Aldosari, M. Molinas, F.A. Alturki, Selecting EEG channels and J. Mariani, K. Kinugawa, Diagnosis of Alzheimer’s disease with
features using multi-objective optimization for accurate MCI detection: validation Electroencephalography in a differential framework, PLoS One 13 (2018)
using leave-one-subject-out strategy, Sci. Rep. 14 (2024) 12483. e0193607.
[34] F. Duan, Z. Huang, Z. Sun, Y. Zhang, Q. Zhao, A. Cichocki, Z. Yang, J. Sole-Casals, [61] Y. Suzuki, M. Suzuki, K. Shigenobu, K. Shinosaki, Y. Aoki, H. Kikuchi, T. Baba,
Topological network analysis of early Alzheimer’s disease based on resting-state M. Hashimoto, T. Araki, K. Johnsen, A prospective multicenter validation study of
EEG, IEEE Trans. Neural Syst. Rehabil. Eng. 28 (2020) 2164–2172. a machine learning algorithm classifier on quantitative electroencephalogram for
[35] H. Wu, L. Kong, Y. Zeng, H. Bao, Resting-state brain connectivity via multivariate differentiating between dementia with Lewy bodies and Alzheimer’s dementia,
EMD in mild cognitive impairment, IEEE Trans. Cogn. Dev. Syst. 14 (2021) PLoS One 17 (2022) e0265484.
552–564. [62] Z. Zavecz, V.D. Shah, O.G. Murillo, R. Vallat, B.A. Mander, J.R. Winer, W.
[36] J. Jiang, Z. Yan, C. Sheng, M. Wang, Q. Guan, Z. Yu, Y. Han, J. Jiang, A novel J. Jagust, M.P. Walker, NREM sleep as a novel protective cognitive reserve factor
detection tool for mild cognitive impairment patients based on eye movement and in the face of Alzheimer’s disease pathology, BMC Med. 21 (2023) 156.
electroencephalogram, J. Alzheimer’s Dis. 72 (2019) 389–399. [63] P.A. Cicalese, R. Li, M.B. Ahmadi, C. Wang, J.T. Francis, S. Selvaraj, P.E. Schulz,
[37] Z.-Z. Cai, R. Lin, X.-X. Wang, Y.-J. Yan, H. Li, Effects of mindfulness in patients Y. Zhang, An EEG-fNIRS hybridization technique in the four-class classification of
with mild cognitive impairment with insomnia: A double-blind randomized alzheimer’s disease, J. Neurosci. Methods 336 (2020) 108618.
controlled trial, Geriatr. Nurs. (Minneap) 47 (2022) 239–246. [64] J. Cruzat, R. Herzog, P. Prado, Y. Sanz-Perl, R. Gonzalez-Gomez, S. Moguilner, M.
[38] D. Stogl, O. Armbruster, M. Mende, B. Hein, X. Wang, P. Meyer, Robot-based L. Kringelbach, G. Deco, E. Tagliazucchi, A. Ibañez, Temporal irreversibility of
training for people with mild cognitive impairment, IEEE Robot. Autom. Lett. 4 large-scale brain dynamics in Alzheimer’s disease, J. Neurosci. 43 (2023)
(2019) 1916–1923. 1643–1656.
[39] H. Zeng, X. Fang, Y. Zhao, J. Wu, M. Li, H. Zheng, F. Xu, D. Pan, G. Dai, EMCI: a [65] L.C. Nunes, P.R. Pinheiro, M.C.D. Pinheiro, M. Simão Filho, R.E.C. Nunes, P.G.C.
novel EEG-based mental workload assessment index of mild cognitive D Pinheiro, Automatic detection and diagnosis of neurologic diseases, IEEE
impairment, IEEE Trans. Biomed. Circuits. Syst. 16 (2022) 902–914. Access 7 (2019) 29924–29941.
[40] X. Bi, H. Wang, Early Alzheimer’s disease diagnosis based on EEG spectral images [66] M. Nour, U. Senturk, K. Polat, A novel hybrid model in the diagnosis and
using deep learning, Neural Netw. 114 (2019) 119–135. classification of Alzheimer’s disease using EEG signals: Deep ensemble learning
[41] C. Porcaro, F. Vecchio, F. Miraglia, G. Zito, P.M. Rossini, Dynamics of the (DEL) approach, Biomed. Signal. Process. Control 89 (2024) 105751.
“cognitive” brain wave P3b at rest for Alzheimer dementia prediction in mild [67] M. Hata, Y. Miyazaki, K. Mori, K. Yoshiyama, S. Akamine, H. Kanemoto, S. Gotoh,
cognitive impairment, Int. J. Neural Syst. 32 (2022) 2250022. H. Omori, A. Hirashima, Y. Satake, Utilizing portable electroencephalography to
[42] E.Q. Wu, X.-Y. Peng, S.-D. Chen, X.-Y. Zhao, Z.-R. Tang, Detecting Alzheimer’s screen for pathology of Alzheimer’s disease: a methodological advancement in
dementia degree, IEEE Trans. Cogn. Dev. Syst. 14 (2020) 116–125. diagnosis of neurodegenerative diseases, Front. Psychiatry 15 (2024) 1392158.
[43] H. Yu, X. Lei, Z. Song, C. Liu, J. Wang, Supervised network-based fuzzy learning [68] M. Buscema, F. Vernieri, G. Massini, F. Scrascia, M. Breda, P.M. Rossini, E. Grossi,
of EEG signals for Alzheimer’s disease identification, IEEE Trans. Fuzzy Syst. 28 An improved I-FAST system for the diagnosis of Alzheimer’s disease from
(2019) 60–71. unprocessed electroencephalograms by using robust invariant features, Artif.
[44] Z. Song, B. Deng, J. Wang, R. Wang, Biomarkers for Alzheimer’s disease defined Intell. Med. 64 (2015) 59–74.
by a novel brain functional network measure, IEEE Trans. Biomed. Eng. 66 (2018) [69] P. Durongbhan, Y. Zhao, L. Chen, P. Zis, M. De Marco, Z.C. Unwin, A. Venneri,
41–49. X. He, S. Li, Y. Zhao, A dementia classification framework using frequency and
[45] B.A. Kent, A.A. Casciola, S.K. Carlucci, M. Chen, S. Stager, M.S. Mirian, P. Slack, time-frequency features based on EEG signals, IEEE Trans. Neural Syst. Rehabil.
J. Valerio, M.J. McKeown, H.H. Feldman, Home EEG sleep assessment shows Eng. 27 (2019) 826–835.
reduced slow-wave sleep in mild–moderate Alzheimer’s disease, Alzheimer’s [70] S.-K. Kim, H. Kim, S.H. Kim, J.B. Kim, L. Kim, Electroencephalography-based
Dementia: Transl. Res. Clin. Interv. 8 (2022) e12347. classification of Alzheimer’s disease spectrum during computer-based cognitive
[46] S. Afshari, M. Jalili, Directed functional networks in Alzheimer’s disease: testing, Sci. Rep. 14 (2024) 5252.
disruption of global and local connectivity measures, IEEE J. Biomed. Health [71] S. Kongwudhikunakorn, S. Kiatthaveephong, K. Thanontip, P. Leelaarporn,
Inform. 21 (2016) 949–955. M. Piriyajitakonkij, T. Charoenpattarawut, P. Autthasan, R. Chaisaen, P. Dujada,
[47] D. Klepl, F. He, M. Wu, D.J. Blackburn, P. Sarrigiannis, EEG-based graph neural T. Sudhawiyangkul, A pilot study on visually stimulated cognitive tasks for EEG-
network classification of Alzheimer’s disease: An empirical evaluation of based dementia recognition, IEEe Trans. Instrum. Meas. 70 (2021) 1–10.
functional connectivity methods, IEEE Trans. Neural Syst. Rehabil. Eng. 30 [72] N. Sharma, M.H. Kolekar, K. Jha, Iterative filtering decomposition based early
(2022) 2651–2660. dementia diagnosis using EEG with cognitive tests, IEEE Trans. Neural Syst.
[48] M. Lopes, R. Cassani, T.H. Falk, Using CNN Saliency Maps and EEG Modulation Rehabil. Eng. 28 (2020) 1890–1898.
Spectra for Improved and More Interpretable Machine Learning-Based [73] H. Azami, S.E. Arnold, S. Sanei, Z. Chang, G. Sapiro, J. Escudero, A.S. Gupta,
Alzheimer’s Disease Diagnosis, Comput. Intell. Neurosci. 2023 (2023) 3198066. Multiscale fluctuation-based dispersion entropy and its applications to
[49] X. Shan, J. Cao, S. Huo, L. Chen, P.G. Sarrigiannis, Y. Zhao, Spatial–temporal neurological diseases, IEEE Access 7 (2019) 68718–68733.
graph convolutional network for Alzheimer classification based on brain [74] Y. Chen, C. Hu, B. Hu, L. Hu, H. Yu, C. Miao, Inferring cognitive wellness from
functional connectivity imaging of electroencephalogram, Hum. Brain Mapp. 43 motor patterns, IEEE Trans. Knowl. Data Eng. 30 (2018) 2340–2353.
(2022) 5194–5209. [75] M.-j. Kim, Y.C. Youn, J Paik, Deep learning-based EEG analysis to classify normal,
[50] M. Imani, Alzheimer’s diseases diagnosis using fusion of high informative BiLSTM mild cognitive impairment, and dementia: Algorithms and dataset, Neuroimage
and CNN features of EEG signal, Biomed. Signal. Process. Control 86 (2023) 272 (2023) 120054.
105298. [76] Y. Pan, M. Liu, Y. Xia, D. Shen, Disease-image-specific learning for diagnosis-
[51] C. Roncero-Parra, A. Parreño-Torres, R. Sánchez-Reolid, J. Mateo-Sotos, A. oriented neuroimage synthesis with incomplete multi-modality data, IEEE Trans.
L. Borja, Inter-hospital moderate and advanced Alzheimer’s disease detection Pattern. Anal. Mach. Intell. 44 (2021) 6839–6853.
through convolutional neural networks, Heliyon (2024) 10. [77] Y. Watanabe, Y. Miyazaki, M. Hata, R. Fukuma, Y. Aoki, H. Kazui, T. Araki,
[52] C.J. Huggins, J. Escudero, M.A. Parra, B. Scally, R. Anghinah, A.V.L. De Araújo, L. D. Taomoto, Y. Satake, T. Suehiro, A deep learning model for the detection of
F. Basile, D. Abasolo, Deep learning of resting-state electroencephalogram signals various dementia and MCI pathologies based on resting-state
for three-class classification of Alzheimer’s disease, mild cognitive impairment electroencephalography data: A retrospective multicentre study, Neural Netw.
and healthy ageing, J. Neural Eng. 18 (2021) 046087. 171 (2024) 242–250.
[53] I.A. Fouad, F.E.-Z.M. Labib, Identification of Alzheimer’s disease from central [78] E. Sibilano, A. Brunetti, D. Buongiorno, M. Lassi, A. Grippo, V. Bessi, S. Micera,
lobe EEG signals utilizing machine learning and residual neural network, Biomed. A. Mazzoni, V. Bevilacqua, An attention-based deep learning approach for the
Signal. Process. Control 86 (2023) 105266. classification of subjective cognitive decline and mild cognitive impairment using
[54] D. kumar Ravikanti, S. Saravanan, EEGAlzheimer’sNet: Development of resting-state EEG, J. Neural Eng. 20 (2023) 016048.
transformer-based attention long short term memory network for detecting [79] H. Jia, Z. Huang, C.F. Caiafa, F. Duan, Y. Zhang, Z. Sun, J. Solé-Casals, Assessing
Alzheimer disease using EEG signal, Biomed. Signal. Process. Control 86 (2023) the Potential of Data Augmentation in EEG Functional Connectivity for Early
105318. Detection of Alzheimer’s Disease, Cognit. Comput. 16 (2024) 229–242.
[55] S.K. Khare, U.R. Acharya, Adazd-Net: Automated adaptive and explainable [80] S. Fouladi, A.A. Safaei, N. Mammone, F. Ghaderi, M.J. Ebadi, Efficient deep
Alzheimer’s disease detection system using EEG signals, Knowl. Based. Syst. 278 neural networks for classification of Alzheimer’s disease and mild cognitive
(2023) 110858. impairment from scalp EEG recordings, Cognit. Comput. 14 (2022) 1247–1268.
[56] Z. Song, B. Deng, J. Wang, G. Yi, An EEG-based systematic explainable detection [81] C. Ieracitano, N. Mammone, A. Bramanti, A. Hussain, F.C. Morabito,
framework for probing and localizing abnormal patterns in Alzheimer’s disease, A convolutional neural network approach for classification of dementia stages
J. Neural Eng. 19 (2022) 036007. based on 2D-spectral representation of EEG recordings, Neurocomputing 323
[57] J. Sekhar, C. Rajyalakshmi, S. Nagaraj, S. Sankar, R. Saturi, A. Harshavardhan, (2019) 96–107.
Deep generative adversarial networks with marine predators algorithm for

17
M. Acharya et al. Computer Methods and Programs in Biomedicine 259 (2025) 108506

[82] M.N.A. Tawhid, S. Siuly, E. Kabir, Y. Li, Exploring frequency band-based is sensitive to early stage Alzheimer’s disease, Clin. Neurophysiol. 113 (2002)
biomarkers of EEG signals for mild cognitive impairment detection, IEEE Trans. 1052–1058.
Neural Syst. Rehabil. Eng. (2023). [100] M. Pietto, M.A. Parra, N. Trujillo, F. Flores, A.M. García, J. Bustin, P. Richly,
[83] H. Zhou, L. Yin, R. Su, Y. Zhang, Y. Yuan, P. Xie, X. Li, STCGRU: A hybrid model F. Manes, F. Lopera, A. Ibáñez, Behavioral and electrophysiological correlates of
based on CNN and BiGRU for mild cognitive impairment diagnosis, Comput. memory binding deficits in patients at different risk levels for Alzheimer’s disease,
Methods Programs Biomed. 248 (2024) 108123. J. Alzheimer’s Dis. 53 (2016) 1325–1340.
[84] Z. You, R. Zeng, X. Lan, H. Ren, Z. You, X. Shi, S. Zhao, Y. Guo, X. Jiang, X. Hu, [101] A. Miltiadous, K.D. Tzimourta, T. Afrantou, P. Ioannidis, N. Grigoriadis, D.
Alzheimer’s disease classification with a cascade neural network, Front. Public G. Tsalikakis, P. Angelidis, M.G. Tsipouras, E. Glavas, N. Giannakeas, A dataset of
Health 8 (2020) 584387. EEG recordings from: Alzheimer’s disease, Frontotemporal dementia and Healthy
[85] A.T. Adebisi, H.-W. Lee, K.C. Veluvolu, EEG-based brain functional network subjects, OpenNeuro 1 (2023) 88.
analysis for differential identification of dementia-related disorders and their [102] Y. Si, R. He, L. Jiang, D. Yao, H. Zhang, P. Xu, X. Ma, L. Yu, F. Li, Differentiating
onset, IEEE Trans. Neural Syst. Rehabil. Eng. (2024). between Alzheimer’s Disease and Frontotemporal Dementia Based on the Resting-
[86] P.M. Rodrigues, B.C. Bispo, C. Garrett, D. Alves, J.P. Teixeira, D. Freitas, State Multilayer EEG Network, IEEE Trans. Neural Syst. Rehabil. Eng. (2023).
Lacsogram: A new EEG tool to diagnose Alzheimer’s disease, IEEE J. Biomed. [103] S. Siuly, Ö.F. Alçin, H. Wang, Y. Li, P. Wen, Exploring rhythms and channels-
Health Inform. 25 (2021) 3384–3395. based EEG biomarkers for early detection of alzheimer’s disease, IEEE Trans.
[87] Y. Shi, Q. Ma, C. Feng, M. Wang, H. Wang, B. Li, J. Fang, S. Ma, X. Guo, T. Li, Emerg. Top. Comput. Intell. (2024).
Microstate feature fusion for distinguishing AD from MCI, Health Inf. Sci. Syst. 10 [104] D. Wen, Z. Bian, Q. Li, L. Wang, C. Lu, X. Li, Resting-state EEG coupling analysis of
(2022) 16. amnestic mild cognitive impairment with type 2 diabetes mellitus by using
[88] B. Wallace, F. Knoefel, R. Goubran, R.A.L. Zunini, Z. Ren, A. Maccosham, EEG/ permutation conditional mutual information, Clin. Neurophysiol. 127 (2016)
ERP: Within episodic assessment framework for cognition, IEEE Trans. Instrum. 335–348.
Meas. 66 (2017) 2525–2534. [105] D. Wen, P. Li, Y. Zhou, Y. Sun, J. Xu, Y. Liu, X. Li, J. Li, Z. Bian, L. Wang, Feature
[89] L. Spyrou, M. Parra, J. Escudero, Complex tensor factorization with PARAFAC2 classification method of resting-state EEG signals from amnestic mild cognitive
for the estimation of brain connectivity from the EEG, IEEE Trans. Neural Syst. impairment with type 2 diabetes mellitus based on multi-view convolutional
Rehabil. Eng. 27 (2018) 1–12. neural network, IEEE Trans. Neural Syst. Rehabil. Eng. 28 (2020) 1702–1709.
[90] S. Chen, C. Zhang, H. Yang, L. Peng, H. Xie, Z. Lv, Z.-G. Hou, A Multi-Modal [106] H.W. Loh, C.P. Ooi, S. Seoni, P.D. Barua, F. Molinari, U.R. Acharya, Application of
Classification Method for Early Diagnosis of Mild Cognitive Impairment and explainable artificial intelligence for healthcare: A systematic review of the last
Alzheimer’s Disease Using Three Paradigms With Various Task Difficulties, IEEE decade (2011–2022), Comput. Methods Programs Biomed. 226 (2022) 107161.
Trans. Neural Syst. Rehabil. Eng. 32 (2024) 1456–1465. [107] M. Abdar, F. Pourpanah, S. Hussain, D. Rezazadegan, L. Liu, M. Ghavamzadeh,
[91] T. Dhar, N. Dey, S. Borra, R.S. Sherratt, Challenges of deep learning in medical P. Fieguth, X. Cao, A. Khosravi, U.R. Acharya, A review of uncertainty
image analysis—improving explainability and trust, IEEE Trans. Technol. Soc. 4 quantification in deep learning: Techniques, applications and challenges, Inf.
(2023) 68–75. Fusion 76 (2021) 243–297.
[92] Q. Teng, Z. Liu, Y. Song, K. Han, Y. Lu, A survey on the interpretability of deep [108] S. Flügge, S. Zimmer, U. Petersohn, arXiv preprint arXiv: 1909.08549, 2019.
learning in medical diagnosis, Multimed. Syst. 28 (2022) 2335–2355. [109] V. Jahmunah, E.Y.K. Ng, R.-S. Tan, S.L. Oh, U.R. Acharya, Uncertainty
[93] A. Atmakuru, S. Chakraborty, O. Faust, M. Salvi, P.D. Barua, F. Molinari, quantification in DenseNet model using myocardial infarction ECG signals,
U. Acharya, N. Homaira, Deep learning in radiology for lung cancer diagnostics: A Comput. Methods Programs Biomed. 229 (2023) 107308.
systematic review of classification, segmentation, and predictive modeling [110] Y. Wang, X. Liu, C. Yu, Assisted diagnosis of alzheimer’s disease based on deep
techniques, Expert. Syst. Appl. (2024) 124665. learning and multimodal feature fusion, Complexity 2021 (2021) 6626728.
[94] D. Klepl, F. He, M. Wu, D.J. Blackburn, P. Sarrigiannis, Adaptive gated graph [111] A.M. Hafiz, S.A. Parah, R.U. Bhat. Attention mechanisms and deep learning for
convolutional network for explainable diagnosis of Alzheimer’s disease using EEG machine vision: A survey of the state of the art, 2021. http://www.doi.org.10
data, IEEE Tran. Neural Syst. Rehabil. Eng. (2023). .21203/rs.3.rs-510910/v1.
[95] M. Kashefpoor, H. Rabbani, M. Barekatain, Automatic diagnosis of mild cognitive [112] S. Georgousis, M.P. Kenning, X. Xie, Graph deep learning: State of the art and
impairment using electroencephalogram spectral features, J. Med. Signals Sens. 6 challenges, IEEe Access 9 (2021) 22106–22140.
(2016) 25–32. [113] P. Liu, X. Qiu, X. Huang, arXiv preprint arXiv: 1609.07222, 2016.
[96] A.M. Pineda, F.M. Ramos, L.E. Betting, A.S. Campanharo, Quantile graphs for [114] P.M. Mammen, arXiv preprint arXiv: 2101.05428, 2021.
EEG-based diagnosis of Alzheimer’s disease, PLoS One 15 (2020) e0231169. [115] B. Yuan, S. Ge, W. Xing, arXiv preprint arXiv: 2005.05083, 2020.
[97] M. Cejnek, O. Vysata, M. Valis, I. Bukovsky, Novelty detection-based approach for [116] H. Zhang, W. Liang, C. Li, Q. Xiong, H. Shi, L. Hu, G. Li, DCML: Deep contrastive
Alzheimer’s disease and mild cognitive impairment diagnosis from EEG, Med. mutual learning for COVID-19 recognition, Biomed. Signal. Process. Control 77
Biol. Eng. Comput. 59 (2021) 2287–2296. (2022) 103770.
[98] E. Gallego-Jutglà, J. Solé-Casals, F.-B. Vialatte, M. Elgendi, A. Cichocki, [117] S. Seoni, A. Shahini, K.M. Meiburger, F. Marzola, G. Rotunno, U.R. Acharya,
J. Dauwels, A hybrid feature selection approach for the early diagnosis of F. Molinari, M. Salvi, All you need is data preparation: A systematic review of
Alzheimer’s disease, J. Neural Eng. 12 (2015) 016018. image harmonization techniques in Multi-center/device studies for medical
[99] T. Musha, T. Asada, F. Yamashita, T. Kinoshita, Z. Chen, H. Matsuda, M. Uno, W. support systems, Comput. Methods Programs Biomed. (2024) 108200.
R. Shankle, A new EEG method for estimating cortical neuronal impairment that

18