2025 International Conference on Intelligent Systems and Computational Networks (ICISCN)

Comprehensive Analysis of Attention

Deficit/Hyperactivity Disorder using Machine
Learning and Deep Learning Algorithms
2025 International Conference on Intelligent Systems and Computational Networks (ICISCN) | 979-8-3315-2924-6/25/$31.00 ©2025 IEEE | DOI: 10.1109/ICISCN64258.2025.10934544

1st T. N. Sindhu 2nd S. R. Sujatha

Department of Computer Science and Engineering Department of Computer Science and Engineering
Sri Siddhartha Institute of Technology Sri Siddhartha Institute of Technology
Tumakuru, India Tumakuru, India
[email protected] [email protected]

Abstract—Attention deficit/hyperactivity disorder (ADHD) [5]. The ADHD diagnosis uses gold standard methods such as
is a neurodevelopmental disorder that affects children from mental status assessments, medical examination,
childhood to adulthood. The symptoms of ADHD include comprehensive interview with children’s caretakers, cognitive
attention deficit, impulsive behaviour, hyperactivity, and early testing, and school reports collected for the analysis of the
diagnosis and treatment are imperative, especially for children patient’s stage to provide better treatment plan and diagnosis
as it continues into adulthood. ADHD detection is generally [6]. Recent clinical studies have focussed on improving
carried out using the existing methods such as Machine techniques for accurate treatment and diagnosis of ADHD [7].
Learning (ML) and Deep Learning (DL) methods. The ML Currently, ADHD diagnosis is carried out using Machine
methods such as Linear Support Vector Machine (SVM), K-
Learning (ML) and Deep Learning (DL) models. The DL
Nearest Neighbors (KNN), Linear Discriminative Analysis
(LDA), Linear Regression (LR), Random Forest (RF), and Least
models help in the diagnosis of ADHD as well as other
Square SVM (LS-SVM) are used for the detection and analysis diseases, but their effectiveness depends on various factors
of ADHD. The existing DL methods such as Deep Convolutional such as the quality of data and the choice of the model [8]. The
Neural Network (DCNN), Graph Recurrent Unit (GRU), ML models simplify complex cases for ADHD diagnosis. The
Convolutional Long Short-Term Memory (ConvLSTM), and diagnostic process of ADHD is hence strived at being
Bidirectional LSTM are also employed for the detection of improved by using ML and DL models [9].
ADHD. The ML and DL algorithms enhance ADHD detection The remainder of the paper is arranged in the following
and diagnosis by automatically determining large patterns in
manner: Section 2 discuss the literature survey, Section 3
data, thereby offering higher accuracy. The performance of the
existing techniques is evaluated in terms of evaluation metrics:
describes the taxonomy for ADHD detection, Section 4
accuracy, precision, sensitivity, and F1-score to assess their represents the comparative analysis of various models, while
effectiveness. Section 5 discuss the problems about ADHD detection, and
finally, Section 6 describes the conclusion of paper.
Keywords—attention-deficit/hyperactivity disorder, deep
learning, impulsive behaviour, machine learning, treatment plans. II. LITERATURE REVIEW
Hasan and Singhal [10] introduced a Support Vector
I. INTRODUCTION Machine (SVM) combined with t-test based methods to
Attention-deficit /Hyperactivity disorder (ADHD) is a determine the optimal channels for Electroencephalogram
neurodevelopmental disorder which affects the brain ability in (EEG) signals for ADHD detection. Then, Least Absolute
children with symptoms including inattention, hyperactivity, Shrinkage and Selection Operator regression (LASSO)
impulsivity, head injuries, and thyroid complications. logistic regression were applied to select significant features
Additionally, the children affected with ADHD are also prone from the chosen channels. The Machine Learning (ML)
to various other conditions such as Major Depressive Disorder classifiers such as Gaussian Process Classification (GPC),
(MDD) and Autism Spectrum Disorder (ASD) [1]. The Random Forest (RF), K-Nearest Neighbors (KNN),
symptoms of ADHD affect children and continue into Multilayer Perceptron (MLP), Decision Tree (DT), and
adulthood, thereby having a long-term impact. Therefore, Logistic Regression (LR) were used to detect ADHD in
early detection, accurate diagnosis, and effective treatment are children. The LASSO with GPC-based system showed better
very crucial in the affected children [2]. ADHD disease performance in identifying children with ADHD from healthy
continues into adulthood and is twice as common in boys than controls. Although the LASSO logistic regression was used
in girls, as boys face greater challenges in adulthood due to for feature selection, it did not consistently capture the most
commitments at home and in the workplace [3]. The causes of important features because of the presence of noise and
ADHD disease include genetic factors, low birth weight, multicollinearity features.
problems in the nervous system, and parental drug usage Shin et al. [11] presented a ML method used to detect
during pregnancy. It also affects the daily behaviours and ADHD in children with Austim Spectrum Disorder (ASD),
activities of children. But, there are no accurate medicines for assuming their hand writing patterns as input. The handwritten
ADHD, and only proper plans and regular treatments can save data sets were used as input by splitting them into two phases:
the lives of these children [4]. training and testing. The ML models of SVM, RF, DT, KNN,
The World Health Organization (WHO) conducted and LR were trained in the training phase, while during the
research on ADHD and found that the number of affected testing phase, the child’s possible ADHD with ASD was
children has significantly increased over the past ten years. predicted. The features were extracted, normalized and
Researchers have been trying to enhance ADHD diagnosis selected by Sequential Forward Floating Selection (SFFS).
with various models and diagnostic methods over the years The RF classifier achieved high accuracy through prediction-

979-8-3315-2924-6/25/$31.00 ©2025 IEEE

Authorized licensed use limited to: Somaiya University. Downloaded on June 18,2025 at 05:22:14 UTC from IEEE Xplore. Restrictions apply.
based patterns which helped extract the relevant features from ASD manage their symptoms more effectively. However, the
the data and enabled the RF classifier predict more accurately CNN-Bi-LSTM model performed well on smaller datasets,
than other classifiers in ML models. However, this ML model but was complex to be applied on larger datasets.
concentrated only on the right hand writing samples, while not
concentrating on the left hand writing samples. Loh et al. [14] developed an EEG-based DL system for the
detection of conduct plan (CD), ADHD+CD, and ADHD. The
Maniruzzaman et al. [12] developed a ML-based model to pre-processing techniques were applied to segment the EEG
identify biomarkers for children with ADHD. A multi-center signals and convert them using a channel-wise Continuous
approach was applied to collect the Reverse Stroop Tasks Wavelet Transform (CWT), which provided both time and
(RST) data comprising of children’s age, and their frequency information. The deep CCN model was used to
behavioural and physical indicators. The brain activity data classify the CWT correlation matrices for the detection of CD,
was measured by variations in blood oxygenation using ADHD, and ADHD+CD. The DL detection system identified
functional near-infrared spectroscopy (fNIRs) and then was particular time segments from the EEG signals and selected
collected to identify potential biomarkers for ADHD. The significant EEG channels for diagnosis, thereby ensuring
training and testing data were then used for identify the aiding professions in diagnosis. However, this model relied on
biomarkers through training the ML models: SVM, KNN, LR, EEG signals which were more complex to pre-process than
Linear Discriminative Analysis (LDA) and RF. The Electrocardiograph (ECG) signals, which presented a setback
combination of these ML models reduced overfitting which of this approach.
facilitated accurate detection. But, the proposed ML-based
model utilized limited data to identify ADHD, and exhibited III. TAXONOMY
computational complexity when used with larger data. ADHD detection involves the analysis of the collected
Shin et al. [13] developed Convolutional Neural Network health data using various ML and DL algorithms, combined
(CNN) combined with Bi-directional Long-Term Memory with ML-based classifiers, namely, Linear SVM and KNN,
(Bi-LSTM) for the detection of ADHD with coexisting ASD and DL-based classifiers namely, Deep Convolutional Neural
in children. The fNIRs signals recorded zigzag and Periodic Network (DCNN), Graph Recurrent Unit (GRU), and
Lines (PL) drawn by children, which were analysed using Convolutional Long Short-Term Memory (ConvLSTM),
statistical analysis. These recorded lines were fed into the which are used to improve the performance of the model.
CNN-Bi-LSTM classifiers to detect ADHD. The early Figure 1 represents the taxonomy of ADHD detection.
detection of these conditions helped children with ADHD and

Fig. 1. Taxonomy of ADHD detection

A. Machine Learning Techniques Mode Network. It was then used to classify these patterns
ML technique is used for classification of new and unseen which represent either the spatial or verbal working memory
data, and to make predictions through training a given dataset. tasks. The SVM classifier effectively handled high-
During the training period, the ML adjusts the model’s dimensional data using fMRI neural patterns by accurately
internal parameters that helps decrease errors in predictions. classifying the complex process.
The commonly used ML methods include SVM and KNN. 2) Support vector machine: Zhang et al. [16] developed a
SVM classifier to classify ADHD based on fingertip force
1) Linear support vector machine: Vilgis et al. [15]
control masks. The SVM classifier was used to analyse the
developed a SVM classifier to identify differences in neural
force data and amplitude resolution, making it suitable for
patterns while performing spatial and verbal working
small datasets. The statistical indicators were derived from
memory tasks in boys with ADHD and Persistent Depressive
reaction time and force fluctuation were selected as input
Disorder (PDD). The SVM classifier was applied to brain
features to develop a data-driven classification algorithm.
imaging data (fMRI - Frequency Magnetic Resonance
The SVM classifier effectively classified children with
Imaging) and trained on neural activation patterns,
ADHD, enabling an objective method for ADHD diagnosis
particularly beta estimates from various brain networks, such
in small sample size and with limited feature dimensions.
as Frontoparietal Network, Salience Network, and Default

Authorized licensed use limited to: Somaiya University. Downloaded on June 18,2025 at 05:22:14 UTC from IEEE Xplore. Restrictions apply.
 3) Least square support vector machine and k-nearest classifier. The Nested Residual Convolutional Denoising
neighbors: Swadi and Miften [17] introduced Least Square Autoencoder (NRCDAE) was used to reduce spatial
SVM (LS-SVM) and KNN to detect ADHD from EEG dimensions and to extract 3D spatial features. The
signals. LS-SVM was used to classify the features, while convolutional GRU served as the classifier employed to
KNN was employed to analyse and select the segmented extract the spatial and temporal features from the reduced
features and remove unnecessary ones. They categorized the dimension data, and to process the spatial features of Resting
EEG signals into five frequency bands using Discrete State- Functional Magnetic Resonance Imaging (rs-fMRI)
Wavelet Transform (DWT). Both these ML models were images. The sigmoid classifier was used to classify the spatio-
utilized to identify ADHD patients and healthy controls, but temporal features of rs-fMRI data. The GRU model
the LS-SVM performed better than the KNN in detecting effectively increased the model’s accuracy of ADHD
ADHD. diagnosis using rs-fMRI images.
3) Convolutional long short-term memory: Bakhtyari and
B. Deep Learning Techniques
Mirzaei [20] developed a combination of ConvLSTM and
DL is a subcategory of AI that was utilized a multi-layered attention mechanism for the classification of ADHD and the
neural network named as Deep Neural Network (DNN) to
control groups. The convolution layer encoded spatial
make decisions, even for complex data. The DL classification
techniques such as CNN, Recurrent Neural Network (RNN), features between channels, while the LSTM captured time
and DCNN have been studied in the existing researches. dependencies, and the attention layer learned from ADHD
classification tasks and converted them to a spatio-temporal
1) Deep convolutional neural network: Sravani et al. [18] representation of EEG recordings. It also detected the
developed a DCNN classifier and Dipper-Throated Particle differences among temporal segments. The attention
Swarm Optimization (DTPSO) were used to detect its mechanism assigned complex weights to crucial temporal
performance of ADHD method. Preprocessing techniques segments, which improved its classification performance.
such as data augmentation and canny deriche edge detection The combination of ConvLSTM and the attention mechanism
were used to discard unclear images and detect the edges. achieved high accuracy for both ADHD patients and the
Further, data augmentation was used to improve the quality healthy group.
of datasets images. The DTPSO algorithm was used to select
the optimized hyper-parameters and create an optimal DCNN IV. COMPARATIVE ANALYSIS
model. The hyperparameter tuning achieved an optimal Table 1 presents a comparative analysis of ADHD
performance and enhanced the accuracy of DCNN detection of the existing approaches using ML and DL
classification. The combination of preprocessing and models. The comparative analysis includes about these
optimization improved the model’s effectiveness of ADHD method’s methodology, advantages, limitations, performance
detection. metrics, and the authors’ names.
2) Graph recurrent unit: Liu et al. [19] developed a Table 1 presents the comparative analysis of ADHD
spatio-temporal feature extraction method and GRU detection.

TABLE I. COMPARATIVE ANALYSIS OF ADHD DETECTION

S. Author Name Methodology Advantages Limitations Performance
No metrics
1. Hasan and Ensemble The LASSO with GPC-based system The LASSO logistic regression faced Accuracy,
Singhal [10] ML showed effective performance in challenges capturing important features because Recall,
identifying and comparing children with of noises and multicollinearity features. Precision,
ADHD and those without it. F1-Score
2. Shin et al. [11] ML based The RF classifier achieved high accuracy Furthermore, this ML model concentrated only Accuracy,
classifiers with predictions based on patterns by about the right hand writing samples, while not Recall,
extracting relevant features from the concentrating on the left hand writing samples. Precision,
data. This enabled the RF classifier to F1-Score
predict accurately than other classifiers
in the ML model.
3. Sravani and DCNN The DTPSO algorithm improved the The hyperparameter optimization led to Accuracy,
Kuppusamy DCNN by adjusting its settings based on overfitting when tested on one dataset. The Specificity,
[18] the model’s performance. This helped model worked well on the training data but not Sensitivity,
the model learn better from MRI images on new, unseen data. F1-Score.
and increased its accuracy of ADHD
detection.
4. Liu et al. [19] GRU The deep spatio-temporal features from The DL models used NRCDAE and GRU, Accuracy,
fMRI were used to classify ADHD by requiring large datasets for resulting in optimal Sensitivity,
capturing the brain activities with high performance. However, collecting data specific Specificity
accuracy to identify complex patterns of to ADHD patients was complex, especially
in brain functions over time. across different age groups and clinical
environments.
5. Bakhtyari and ConvLSTM The ConvLSTM model enhanced The ConvLSTM faced difficulties with EEG Accuracy,
Mirzaei [20] ADHD analysis by combining signals which caused variations in the brain’s Recall,
convolutional layers with LSTM units, activities and functions, providing inconsistent Precision and
allowing it to capture spatial patterns and results. F1-Score
temporal dependencies in the data.

Authorized licensed use limited to: Somaiya University. Downloaded on June 18,2025 at 05:22:14 UTC from IEEE Xplore. Restrictions apply.
 V. PROBLEM STATEMENT [6] B. Kim, J. Park, T. Kim, and Y. Kwon, “Finding essential parts of the
brain in rs-fMRI can improve ADHD diagnosis using deep learning,”
The detection of ADHD using the existing approaches IEEE Access, vol. 11, pp. 116065-116075, October 2023.
demonstrating certain setbacks during analysis using the ML [7] M. Maniruzzaman, J. Shin, and M. A. M. Hasan, “Predicting children
and DL methods. These challenges are noted as below: with ADHD using behavioral activity: a machine learning analysis,”
Appl. Sci, vol. 12, p. 2737, March 2022.
• ML proved capable in ADHD detection but faced [8] N. Luo, X. Luo, S. Zheng, D. Yao, M. Zhao, Y. Cui, Y. Zhu, V. D.
problems arose with the interpretability and Calhoun, L. Sun, and J. Sui, “Aberrant brain dynamics and spectral
generalization of the model. This made additional power in children with ADHD and its subtypes,” European Child &
precautions such as effective feature selection, Adolescent Psychiatry, vol. 32, pp. 2223-2234, August 2023.
balanced datasets, and avoiding of overfitting, crucial [9] W. Lee, D. Lee, S. Lee, K. Jun, and M. S. Kim, “Deep-learning-based
ADHD classification using children’s skeleton data acquired through
for designing robust models. the ADHD screening game,” Sens, vol. 23, p. 246, December 2023.
• The models faced difficulties with imbalanced and [10] U. Hassan, and A. Singhal, “Convolutional neural network framework
limited datasets during ADHD detection, making the for EEG-based ADHD diagnosis in children,” Health Inf. Sci. Syst, vol.
12, p. 44, August 2024.
ML models hard to train. Moreover, the smaller
[11] J. Shin, M. Maniruzzaman, Y. Uchida, M. A. M. Hasan, A. Megumi,
datasets caused overfitting, thereby rendering a poor and A. Yasumura, “Handwriting-based ADHD detection for children
performance. having ASD using machine learning approaches,” IEEE Access, vol.
11, pp. 84974-84984, August 2023.
• The noise in data concealed the important patterns, [12] M. Maniruzzaman, K. Hirooka, Y. Tomioka, M. A. M. Hasan, Y. S.
making it hard to identify the necessary features and Hwang, A. Megumi, A. Yasumura, and J. Shin, “Machine Learning-
noises from different places which include the Based ADHD Detection From fNIRs Signal During Reverse Stroop
background sounds or movements during data Tasks,” IEEE Access, vol. 12, pp. 82984-82995, June 2024.
collection, posing significant challenges during feature [13] J. Shin, S. Konnai, M. Maniruzzaman, M. A. M. Hasan, K. Hirooka, A.
extraction. Megumi, and A. Yasumura, “Identifying ADHD for children with
coexisting ASD from fNIRs signals using deep learning approach,”
VI. CONCLUSION IEEE Access, vol. 11, pp. 82794-82801, July 2023.
[14] H. W. Loh, C. P. Ooi, S. L. Oh, P. D. Barua, Y. R. Tan, U. R. Acharya,
ADHD is a neurodevelopmental disorder which affects and D. S. S. Fung, “ADHD/CD-NET: automated EEG-based
people from childhood into adulthood. ADHD is a long-term characterization of ADHD and CD using explainable deep neural
condition and hence requires early detection, accurate network technique,” Cognit. Neurodyn, vol. 18, pp. 1609-1625,
diagnosis, and effective treatment for affected the children. November 2024.
The ADHD is influenced factors such as genetics, low birth [15] V. Vilgis, D. Yee, T. J. Silk, and A. Vance, “Distinct neural profiles of
frontoparietal networks in boys with ADHD and boys with persistent
weight, problems in nervous system, and substance usage depressive disorder,” Cognitive, Affective, & behavioral
during pregnancy. Many existing researches and classification neuroscience, vol. 22, pp. 1183-1198, March 2022.
techniques have shown exceptional performances in ADHD [16] Z. Zhang, S. Kang, J. Yu, H. Li, G. Yin, H. Zhang, L. Sun, and D.
detection. The ML methods such as SVM, KNN, RF, LDA, Wang, “Quantitative Identification of ADHD Tendency in Children
LR classification techniques, and DL methods such as CNN, With Immersive Fingertip Force Control Tasks,” IEEE Trans. Neural
Bi-LSTM, DCNN, ConvLSTM and GRU classification Syst. Rehabil. Eng, vol. 31, pp. 4561-4569, November 2023.
techniques, are currently being studied in researches. The ML [17] A. T. Swadi, and F.S Miften, “Detection of attention deficit
hyperactivity disorder based on EEG signals using Least Square
techniques were mostly chosen to detect ADHD because they Support Vector Machine (LS-SVM),” Computer Methods in
show better performance when compared to DL techniques. Biomechanics and Biomedical Engineering: Imaging &
The existing model evaluate performance in terms of metrics Visualization, vol. 11, pp. 2495-2507, August 2023.
such as accuracy, precision, F1- Score, sensitivity, and [18] K. Sravani, and P. Kuppusamy, “Optimized Deep Convolutional
specificity. However, to enhance ADHD detection, both ML Neural Network for Autism Spectrum Disorder Detection Using
and DL techniques need to consistently advance their Structural MRI and DTPSO,” IEEE Access, vol. 12, pp. 110035-
110051, August 2024.
techniques and algorithms to improve ADHD detection.
[19] S. Liu, L. Zhao, J. Zhao, B. Li, and S. H. Wang, “Attention
REFERENCES deficit/hyperactivity disorder classification based on deep spatio-
temporal features of functional magnetic resonance imaging,” Biomed.
[1] K. F. Zhang, S. C. Yeh, E. H. K. Wu, X. Xu, H. J. Tsai, and C. C. Chen, Signal Process. Control, vol. 71, Part B, p. 103239, January 2022.
“Fusion of Multi-Task Neurophysiological Data to Enhance the [20] M. Bakhtyari, and S. Mirzaei, “ADHD detection using dynamic
Detection of Attention-Deficit/Hyperactivity Disorder,” IEEE J. connectivity patterns of EEG data and ConvLSTM with attention
Transl. Eng. Health Med, vol. 12, pp. 668-674, July 2024. framework,” Biomed. Signal Process. Control, vol. 76, p. 103708, July
[2] K. E. Thornton, K. K. Wiggs, J. N. Epstein, L. Tamm, and S. P. Becker, 2022.
“ADHD and cognitive disengagement syndrome symptoms related to
self-injurious thoughts and behaviors in early adolescents,” Europian
Child Adolesc Psychiatry, September 2024.
[3] M. Cao, E. Martin, and X. Li, “Machine learning in attention-
deficit/hyperactivity disorder: new approaches toward understanding
the neural mechanisms,” Transl. Psychiatry, vol. 13, p. 236, July 2023.
[4] C. Nash, R. Nair, and S. M. Naqvi, “Insights into Detecting Adult
ADHD Symptoms Through Advanced Dual-Stream Machine
Learning,” IEEE Trans. Neural Syst. Rehabil. Eng, vol. 32, pp. 3378-
3387, August 2024.
[5] H. W. Loh, C. P. Ooi, S. L. Oh, P. D. Barua, Y. R. Tan, F. Molinari, S.
March, U. R. Acharya, and D. S. S. Fung, “Deep neural network
technique for automated detection of ADHD and CD using ECG
signal,” Comput. Methods Programs Biomed, vol. 241, p. 107775,
November 2023.

Authorized licensed use limited to: Somaiya University. Downloaded on June 18,2025 at 05:22:14 UTC from IEEE Xplore. Restrictions apply.