An ADHD Diagnostic Interface Based on EEG Spectrograms and Deep Learning Techniques

Abstract

This paper introduces an innovative approach to Attention-deficit/hyperactivity disorder (ADHD) diagnosis by employing deep learning (DL) techniques on electroencephalography (EEG) signals. This method addresses the limitations of current behavior-based diagnostic methods, which often lead to misdiagnosis and gender bias. By utilizing a publicly available EEG dataset and converting the signals into spectrograms, a Resnet-18 convolutional neural network (CNN) architecture was used to extract features for ADHD classification. The model achieved a high precision, recall, and an overall F1 score of 0.9. Feature extraction highlighted significant brain regions (frontopolar, parietal, and occipital lobes) associated with ADHD. These insights guided the creation of a three-part digital diagnostic system, facilitating cost-effective and accessible ADHD screening, especially in school environments. This system enables earlier and more accurate identification of students at risk for ADHD, providing timely support to enhance their developmental outcomes. This study showcases the potential of integrating EEG analysis with DL to enhance ADHD diagnostics, presenting a viable alternative to traditional methods.

Full text

An ADHD Diagnostic Interface Based on EEG
Spectrograms and Deep Learning Techniques
Medha Pappula
Children’s National Hospital
Washington, DC, USA
mpappula@childrensnational.org
Syed Muhammed Anwar
Children’s National Hospital
Washington, DC, USA
sanwar@childrensnational.org
Abstract—This paper introduces an innovative approach to
Attention-deficit/hyperactivity disorder (ADHD) diagnosis by em-
ploying deep learning (DL) techniques on electroencephalography
(EEG) signals. This method addresses the limitations of current
behavior-based diagnostic methods, which often lead to misdi-
agnosis and gender bias. By utilizing a publicly available EEG
dataset and converting the signals into spectrograms, a Resnet-18
convolutional neural network (CNN) architecture was used to ex-
tract features for ADHD classification. The model achieved a high
precision, recall, and an overall F1 score of 0.9. Feature extraction
highlighted significant brain regions (frontopolar, parietal, and
occipital lobes) associated with ADHD. These insights guided
the creation of a three-part digital diagnostic system, facilitating
cost-effective and accessible ADHD screening, especially in school
environments. This system enables earlier and more accurate
identification of students at risk for ADHD, providing timely
support to enhance their developmental outcomes. This study
showcases the potential of integrating EEG analysis with DL to
enhance ADHD diagnostics, presenting a viable alternative to
traditional methods.
Index Terms—ADHD, diagnosis, Pediatric, Assessment, Early
detection, Electroencephalography, Deep Learning
I. INTRODUCTION
Attention-deficit/hyperactivity disorder (ADHD) is a
widespread neuro-developmental disorder, estimated to affect
about 10% of the global population [1]. Characterized by
challenges in attention control, ADHD interferes with many
aspects of daily life functioning and long-term development,
manifesting in three predominant subtypes: hyperactive,
inattentive, and both combined [2]. Understanding these
variations is crucial for an accurate diagnosis and targeted
interventions.
The current diagnostic approach, Diagnostic and Statistical
Manual of Mental Disorders, Fifth Edition (DSM-V) relies
heavily on behavioral observation, focusing on symptoms
present before the age of 12, their interference with various
aspects of life, and the exclusion of other explanations for the
behavior [3,4]. However, this method has limitations, raising
misdiagnosis concerns. This is particularly evident in the
late diagnosis of girls, who often exhibit ADHD symptoms
in a less hyperactive, more inattentive manner. Boys are
more likely to be diagnosed early due to their hyperactivity,
highlighting a gender bias in the diagnostic process [5].
Recent advances in deep learning (DL) methodologies have
opened doors to more data-driven forms of assessments [6-
8]. Physiological signal analysis offers exciting prospects for
more accurate ADHD assessments. One such physiological
signal, electroencephalography (EEG), which measures elec-
trical brain activity as neurons fire, has shown promise given
the neurodevelopmental nature of ADHD [9]. With the rise
of commercial EEG headsets, these models can be used in
tandem with digital screening tools. These, especially in school
environments, can provide low-cost and accessible screening
allowing for students to receive the necessary support earlier
[10].
This research harnesses the power of EEG signals, im-
age processing, and deep learning techniques to develop an
advanced ADHD assessment tool. This model is then used
to develop a novel screening software targeting the most
impacted brain areas identified by the model. This screening
software, pairing with commercial EEG headsets, ultimately
provides a cost-effective method of accurate screening. The
data utilized in this study is taken from a public repository
representing EEG data from children having ADHD and
control group [11]. The data comprises EEG recordings of
variable length from 61 children with ADHD and 60 healthy
controls aged 7-12 undergoing visual attention tasks. These
signals are then transformed into spectrograms which are
fed into a Convolutional Neural Network (CNN), specifically
the Resnet-18 architecture [12]. This serves as the base for
feature extraction. The major contribution is the use of these
features to develop a novel screening system that can be easily
implemented in public settings such as schools.
The rest of the paper is organized as follows. Section II de-
scribes the methodology used, Section III shows experimental
results, and Section IV concludes the work done, presenting
next avenues for research.
II. PRO PO SE D METHODOLOGY
The implementation of this project involved four key sec-
tions shown in Figure 1. The first step was to collect data from
the EEG data for ADHD / control children dataset [11]. The
core of the methodology includes the training and subsequent
feature extraction of the model, utilizing spectrograms to
accurately visualize the data and using CNNs, specifically
arXiv:2412.02695v1 [cs.CY] 3 Dec 2024

Fig. 1. Our proposed approach for ADHD detection using features extracted from EEG spectrograms and deep learning model.
Resnet-18 architecture to develop the classification model.
Resnet-18 in particular was used for its ability to accurately
interpret the complex patterns hidden within the spectrogram
data. The vital step includes taking the results from the feature
extraction to develop a digital assessment program. These steps
result in a robust system able to accurately assess ADHD
in student population. Below is a detailed overview of each
section
A. Electroencephalography (EEG) data preparation
The EEG data for ADHD / control children comprises of
EEG signals of various lengths from 61 children with ADHD
and 60 children without (control), with a total of 121 data
points. This EEG data comprised 19 channels (Fz, Cz, Pz,
C3, C4, T3, T4, Fp1, Fp2, F3, F4, F7, F8, P3, P4, T5, T6,
O1, O2) and was recorded at a frequency of 128 Hz using the
10-20 electrode system. The children in the study were aged
7-12, which is around the recommended age for diagnosis of
ADHD in the DSM-V criteria. Those in the control group
had no previous history of any other psychiatric disorders,
epilepsy, or high-risk behaviors. Each child was subjected to
a visual attention task, counting characters on a screen, which
is a known area of weakness of those with ADHD [13]. EEG
recordings varied in length depending on the response time of
each child. All recordings were stored in raw format using a
.mat file.
To prepare the data, each of the input .mat file was metic-
ulously processed to generate sub-videos. This preparation
involved leveraging Python’s MNE and pandas packages. Ini-
tially, the EEG data was imported as a pandas DataFrame, with
recordings from each individual segmented into distinct sub-
groups. Following this, an MNE info object was constructed to
encapsulate the recording conditions, which were then updated
to reflect the accurate experimental settings. The data was
subsequently converted into an MNE RawArray, allowing for
advanced processing. To enhance signal quality and ensure
higher accuracy, a band-pass filter was applied, spanning from
1 to 30 Hz. The filtered data was then segmented into 3-second
intervals with a 2-second overlap (1-second overlap on either
side), optimizing the temporal resolution for further analysis.
This pre-processing ensures that the data is well-structured and
ready for subsequent analytical steps.
B. Spectrogram generation
These overlapping segments were then separately converted
into spectrograms. In particular the Continuous Wavelet Trans-
form (CWT) is a powerful tool for analyzing signals with time-
frequency resolution, allowing for a deep examination of how
the frequency content of a signal evolves over time. The CWT
involves convolving the input signal with a family of wavelet
functions, each of which is scaled and translated to capture
different frequency components.
CW Tx(a, b) = Z∞
−∞
x(t)·ψ∗t−b
adt (1)
This process is mathematically expressed in Eq. 1 where x(t)
represents the input signal, ψ(t) denotes the wavelet function,
a is the scale parameter, and b is the translation parameter.
In this study, CWT was employed to generate time-frequency
representations of EEG signals. The resulting time-frequency
maps are of fixed size, which influences the input size of the
deep learning model. Fig. 2 shows the spectrograms of binary
classified ADHD diagnosis of different subjects. Such an
approach for EEG analysis has also been found useful in other
tasks such as seizure [14] and motor imagery classification
[15]. Utilizing advanced classification algorithms can help in
learning identifying specific differences between the condition
and control groups.
C. Deep Learning Model
To extract features from the spectrogram data, a CNN
architecture was used. Specifically, this study utilized Resnet-
18, which is well known for its ability to capture intricate
differences in image data. This model, trained from scratch on
the transformed EEG data performed a binary classification on
the individual’s ADHD assessment. Each of the total 43153
samples had a structure of 100 time points over 19 channels.

Fig. 2. Samples of CWT for children with (left) and without (right) ADHD.
A 5-fold cross-validation accuracy was used to evaluate the
model’s effectiveness. Following this initial training, hyper-
parameter tuning for both the model and the spectrogram
generation further enhanced model ability to perform opti-
mally. The model was used as a baseline as a method of
diagnosis. Primarily, the second half of this paper describes
its use for feature extraction rather than binary classification.
This extraction enables the creation of the front-end interface.
D. Feature Extraction
To aid in the development of a non-EEG based screening
device, the most important features were extracted from the
final model. This was done by comparing the baseline and
permuted accuracy. Permuted accuracy was determined by
replacing a specified feature in the testing data with ran-
domized values and recalculating the model’s accuracy. This
method demonstrated how significantly the model relied on
that feature’s data for its predictions. This process was repeated
19 times for each input channel to ensure robust results. As
shown in Fig. 3, electrodes FP1/2 (frontal), P3/4 and P7/8
(parietal), and O1/2 (occipital) are found to be significant in
ADHD assessment
Fig. 3. Relative importance of each EEG channel in overall ADHD prediction.
III. EXPERIMENTAL RES ULTS
A publicly available EEG data for ADHD / control children
dataset was used for the binary classification of ADHD preva-
lence. These recordings were then cleaned and computed into
spectrograms. This computation and the training of the deep
learning model was performed using the Google’s Colabora-
tory platform. Table 1 presents the precision, recall, and F1-
score for the classification with/without ADHD. The results
show that Resnet-18 has strong performance given its high
precision and recall values, along with a consistent high F1
score. This suggests that Resnet-18 effectively captured the
ADHD classification based on EEG data. This 90% accuracy is
similar and at times exceeds other similar studies. For example
86.7% in looking at objective measures and 68.8% training a
model on medical records.
TABLE I
PERFORMANCE METRICS FOR ADHD CL ASS IFIC ATIO N.
Precision Recall F1-Score
0 (No ADHD) 0.98 0.79 0.88
1 (ADHD) 0.86 0.99 0.92
Accuracy - - 0.90
Macro avg 0.92 0.89 0.90
Weighted avg 0.91 0.90 0.90
A. Significant EEG channels
Feature extraction highlights that channels FP1/2 (fron-
topolar), P3/4 (parietal), P7/8 (parietal), and O1/2 (occipital)
are significantly affected in children with ADHD. Changes
or damage in these areas are associated with difficulties in
focusing on multiple stimuli, making decisions, and perceiving
stimuli, which are typical behaviors observed in children
with ADHD [16]. This supports current research suggesting
decreased gray matter in those regions of the brain [17].
B. Front-end interface
The results from feature extraction informed the develop-
ment of a targeted cognitive test system shown in Fig. 4. This
system comprises three tests:
Test 1: Frontopolar Lobe Function- Patients identify
whether two circles are the same or different colors.
Test 2: Parietal Lobe Function- Patients determine the
orientation of a line, using a reference orientation map with
numbered lines. This assesses spatial awareness.

Fig. 4. Visual representation of the ADHD identification tests.
Test 3: Occipital Lobe Function- Patients match an image
with a word, testing the occipital lobe’s ability to retrieve
information.
The reaction time and accuracy can be collected to better
inform the degree of the individual’s behavior, and can be
improved when paired with a commercial EEG headset. This
type of interface is easily implementable in school systems
allowing for more students to be accurately diagnosed earlier.
IV. CONCLUSION
This paper explored the potential use of EEG in ADHD
assessment, employing the Resnet-18 architecture on a pub-
licly available dataset. Significant features were extracted by
applying CNNs on spectrograms developed from EEG activity.
For children with ADHD, the model identified frontopolar,
parietal, and occipital areas to be the most impacted. The
model had an outstanding performance, with a high F1 score of
0.9. The cognitive testing system based on the identified areas
provide a straightforward evaluation, compared to the previous
behavior based methods. Overall, this paper demonstrates the
feasibility EEG use in ADHD diagnosis and introduces a
novel approach to ADHD screening, particularly in settings
like schools. In the future we plan to accurately test this front
end interface and improve upon it, with the hope to deploy it
within public settings.
REFERENCES
[1] “Products - Data Briefs - Number 499 - March 2024,” www.cdc.gov,
Mar. 15, 2024. https://www.cdc.gov/nchs/products/databriefs/db499.htm
[2] A. Singh, C. J. Yeh, N. Verma, and A. K. Das, “Overview
of attention deficit hyperactivity disorder in young children,”
Health Psychology Research, vol. 3, no. 2, Sep. 2015, doi:
https://doi.org/10.4081/hpr.2015.2115.
[3] American Psychiatric Association, 2000. Diagnostic and statistical man-
ual of mental disorders. Text revision.
[4] E. Abdelnour, M. O. Jansen, and J. A. Gold, “ADHD Diag-
nostic Trends: Increased Recognition or Overdiagnosis?,” Missouri
Medicine, vol. 119, no. 5, pp. 467–473, Sep. 2022, Available:
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9616454/
[5] O. Slobodin and M. Davidovitch, “Gender Differences in Objective
and Subjective Measures of ADHD Among Clinic-Referred Children,”
Frontiers in Human Neuroscience, vol. 13, no. 441, Dec. 2019, doi:
https://doi.org/10.3389/fnhum.2019.00441.
[6] RaviPrakash, H. and Anwar, S.M., 2023. Do It Yourself: Wearable
Sensors and AI for Self-Assessment of Mental Health. AI in Clinical
Medicine: A Practical Guide for Healthcare Professionals, pp.94-103.
[7] Arsalan, A., Anwar, S.M. and Majid, M., 2022. Mental stress detection
using data from wearable and non-wearable sensors: a review. arXiv
preprint arXiv:2202.03033.
[8] Arsalan, A., Majid, M. and Anwar, S.M., 2020. Electroencephalography
based machine learning framework for anxiety classification. In Intelli-
gent Technologies and Applications: Second International Conference,
INTAP 2019, Bahawalpur, Pakistan, November 6–8, 2019, Revised
Selected Papers 2 (pp. 187-197). Springer Singapore.
[9] A. Lenartowicz and S. K. Loo, “Use of EEG to Diagnose ADHD,”
Current Psychiatry Reports, vol. 16, no. 11, Sep. 2014, doi:
https://doi.org/10.1007/s11920-014-0498-0.
[10] Ana-Marta Gabald´
on-P´
erez, Mar´
ıa-Luisa Mart´
ın-Ruiz, F. D´
ıez-Mu˜
noz,
Mar´
ıa Dol´
on-Poza, Nuria M´
aximo-Bocanegra, and I. Pau, “The Potential
of Digital Screening Tools for Childhood ADHD in School Environ-
ments: A Preliminary Study,” Healthcare, vol. 11, no. 20, pp. 2795–2795,
Oct. 2023, doi: https://doi.org/10.3390/healthcare11202795.
[11] A. M. Nasrabadi, “EEG data for ADHD / Control children,” ieee-
dataport.org, Jun. 10, 2020. https://ieee-dataport.org/open-access/eeg-
data-adhd-control-children
[12] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning
for Image Recognition,” 2016 IEEE Conference on Computer Vi-
sion and Pattern Recognition (CVPR), pp. 770–778, Jun. 2016, doi:
https://doi.org/10.1109/cvpr.2016.90.
[13] D. Canu et al., “Visual search in neurodevelopmental disorders: evidence
towards a continuum of impairment,” European Child & Adolescent
Psychiatry, Mar. 2021, doi: https://doi.org/10.1007/s00787-021-01756-
z.
[14] Lam, V., Oliugbo, C., Parida, A., Linguraru, M.G. and Anwar, S.M.,
2024, April. Self-supervised learning for seizure classification using
ECoG spectrograms. In Medical Imaging 2024: Computer-Aided Di-
agnosis (Vol. 12927, pp. 606-611). SPIE.
[15] Khan, S.U., Majid, M., Linguraru, M.G. and Anwar, S.M., 2023,
July. Upper Limb Movement Execution Classification using Electroen-
cephalography for Brain Computer Interface. In 2023 45th Annual In-
ternational Conference of the IEEE Engineering in Medicine & Biology
Society (EMBC) (pp. 1-4). IEEE.
[16] Emser, Theresa S., et al. “Assessing ADHD Symptoms in Children
and Adults: Evaluating the Role of Objective Measures - Behavioral
and Brain Functions.” BioMed Central, BioMed Central, 18 May 2018,
doi.org/10.1186/s12993-018-0143-x.
[17] Mikolas, Pavol, et al. “Training a Machine Learning Classifier to
Identify ADHD Based on Real-World Clinical Data from Medical
Records.” Nature News, Nature Publishing Group, 28 July 2022,
www.nature.com/articles/s41598-022-17126-x.
[18] L. Sørensen, S. Adolfsdottir, E. Kvadsheim, H. Eichele, K. J.
Plessen, and E. Sonuga-Barke, “Suboptimal decision making and in-
terpersonal problems in ADHD: longitudinal evidence from a lab-
oratory task,” Scientific Reports, vol. 14, no. 1, Mar. 2024, doi:
https://doi.org/10.1038/s41598-024-57041-x.
[19] R. Baboli, M. Cao, E. Martin, J. M. Halperin, K. Wu, and
X. Li, “Distinct structural brain network properties in children
with familial versus non-familial attention-deficit/hyperactivity dis-
order (ADHD),” Cortex, vol. 179, pp. 1–13, Oct. 2024, doi:
https://doi.org/10.1016/j.cortex.2024.06.019.