Abstract
Attention deficit hyperactivity disorder (ADHD) is a neurobiological condition that appears during an individual's childhood and may follow her/him for life. The research objective was to understand better how and which computer technologies have been applied to support ADHD diagnosis and treatment. The research used the systematic literature review method: a rigorous, verifiable, and repeatable approach that follows well-defined steps. Six well-known academic data sources have been consulted, including search engines and bibliographic databases, from technology and health care areas. After a rigorous research protocol, 1,239 articles were analyzed. For the diagnosis, the use of machine learning techniques was verified in 61 percent of the articles. Neurofeedback was ranked second with 9.3 percent participation, followed by serious games and eye tracking with 5.6 percent each. For the treatment, neurofeedback was present in 50 percent of the articles, whereas some studies combined both approaches, accounting for 31 percent of the total. Nine percent of the articles reported remote assistance technology, whereas another 9 percent have used virtual reality. By highlighting the leading computer technologies used, their applications, results, and challenges, this literature review breaks ground for further investigations. Moreover, the study highlighted the lack of consensus on ADHD biomarkers. The approaches using machine learning call attention to the probable occurrence of overfitting in several studies, thus demonstrating limitations of this technology on small-sized bases. This research also presented the convergence of evidence from different studies on the persistence of long-term effects of using neurofeedback in treating ADHD.
Introduction
Attention deficit hyperactivity disorder (ADHD) is a neurobiological condition that appears during an individual's childhood and may follow her/him for life. Faraone et al. found out that, in ∼65 percent of the cases, ADHD persists into adolescence and adulthood. 1
Even though it is not a new disorder, ADHD diagnosis is complicated to achieve and requires a multidisciplinary team. According to Souza et al., the dimensional and noncategorical nature of psychiatry diagnoses of childhood and adolescence demands new tools and approaches. 2 Moreover, despite the current advanced research stage, ADHD diagnosis remains clinical and subjective, without material biomarkers to guide diagnosis confirmation.
Some researchers claim that it is essential to provide timely and accurate information on the disorder and strategies to handle it, mainly because of common misconceptions about ADHD. 3 In this regard, the treatment phase of the neurological disorder also offers a vast field for further research. Computer technologies have been used for such purposes since 1985, using CAPTAIN. 4 Since then, new technologies have been developed and keep being improved.
The purpose of the research was to investigate how and which computer technologies have been applied to support ADHD diagnosis and treatment. Through a systematic literature review (SLR), it was possible to explore the trends, limits, and existing challenges in the current technological scenario.
Methods
An SLR was carried out, contextualizing the theme in the current scientific scenario. According to Kitchenham, 5 the SLR is a rigorous, verifiable, and repeatable approach that must follow well-defined phases, mainly composed of planning, executing, and publishing results.
Planning phase
The following research questions have been formulated:
RQ1: Which computer technologies are being used in ADHD diagnosis? RQ2: Which computer technologies are being used in ADHD treatment? RQ3: What hit rates have been measured in the diagnosis support surveys? RQ4: What is the sample size used in the surveys? RQ5: What is the impact of computer technology in the treatment phase?
Six databases have been selected: IEEE, PsycINFO, PubMed, ScienceDirect, Scopus, and Web of Science. As part of the SLR protocol, the definition of the search string was a very insightful step. Table 1 relates themes versus search expressions.
Relationship Between Themes and Search Expressions
ADHD, attention deficit hyperactivity disorder.
The following selection and quality criteria were established:
• Inclusion criteria:
CI1: Articles that match the search string used.
CI2: Articles written in English.
CI3: Articles published between 2014 and 2019.
CI4: Articles or publications from journals or conferences.
• Exclusion criteria:
CE1: Duplicate publications.
CE2: Secondary or tertiary studies.
CE3: Articles not addressing ADHD in particular.
CE4: Articles addressing ADHD but not diagnosis or treatment.
CE5: Articles not addressing the use of computer technologies for ADHD diagnosis and treatment.
CE6: Incomplete studies (solutions proposed yet to be applied in research).
• Quality criteria:
CQ1: Journals with impact factor in the first quadrant (Q1) of Scimago Journal Rank (SJR). Journals without impact factor reported by the SJR were not discarded.
CQ2: Each step of protocol development was reviewed by a research group comprising senior researchers, doctoral students, and master's students.
Execution phase
According to Felizardo et al., the execution phase of the SLR comprises the identification and selection stages of the primary studies, through the execution of the search strategy previously defined. 6 The identification step took place by matching the search string to the other five research bases, available in Supplementary Appendix Table SA1. The selection process comprised four phases:
Removal of duplicate publications.
Reading of title, abstract, and keywords.
Application of the quality criteria.
Full reading of the articles.
Table 2 shows the quantities resulting from each step.
Results of Protocol Phases
Results
The 86 finalist articles are listed in Supplementary Appendix Table SA2. A significant number of articles focused on the diagnosis of the disorder (62.8 percent). As for the sample size used, which corresponds to the research question RQ4, the samples presented a considerable dispersion in their distribution. In studies on treatment, the most significant sample used had 552 individuals, and the smallest only 3. On diagnosis, the largest number of respondents was 1,123 individuals, whereas in the smallest case, it was only 10.
Discussion
Computer technology for treatment
Neurofeedback technique is present in half of the articles. Serious games as a means to assist the treatment are ranked second, with 47 percent. Some studies combined the use of both approaches, accounting for 31 percent of the total. Nine percent of the articles reported remote assistance technology, whereas another 9 percent have used virtual reality. Therefore, the research question RQ2, on which computer technologies are being used in the ADHD treatment, could be answered.
Regarding the research question RQ5, about the impact of computational technology in supporting treatment, it was possible to verify that several studies demonstrate the lasting effect of neurofeedback. In some cases, such a result was seen even after a few months later.7–9 Table 3 summarizes the technologies used in the ADHD treatment.
Technologies Used in the Treatment
BCI, brain-computer interface; EEG, electroencephalogram; RMT, remote monitoring technology; rtfMRI-NF, real-time functional magnetic resonance imaging neurofeedback; SMS, short message service; SVM, support vector machine.
Neurofeedback
Cowley et al. chose Enobio as the electroencephalogram (EEG) device, and the software was developed using OpenViBE framework. 10 The technique improved self-reported ADHD symptoms. Other researchers analyzed the differences between ADHD subtype groups using the BrainMaster software. Results indicate that theta/beta protocol was more effective in the predominantly inattentive subset of individuals. 11
Azman et al. evaluated a physical race car game to treat ADHD. Wireless data transmission was used to move the car through brain signals. 12 They combined NeuroSky MindWave Mobile, Arduino UNO, and Toys“R”Us Fastlane Slot Car. As the EEG receives user's brain signal, the car is activated depending on user's concentration. Alternatively, Shin et al. used a tablet to experience situations where neurofeedback could help the executive function. 8 As a result, there was an improvement in cognitive function, even during the follow-up period.
Georgiou et al. 13 developed a serious game called REEFOCUS, which also used augmented reality and neurofeedback. REEFOCUS actively involves parents, doctors, and educators, allowing them to monitor children's progress and adjust the intervention program remotely. The results showed that the use of the game had a greater effect on improving attention.
Other researchers have also identified the persistence of long-term results. Alegria et al. performed tests using real-time functional magnetic resonance imaging (fMRI) neurofeedback. 9 Results showed a reduction in symptoms during the 11-month follow-up period. Another research used the Play Attention system through brain waves sensors built in a bicycle helmet. Individuals maintained the improvements during the 6-month follow-up period. 7 In contrast, Bink et al. reported an experiment with negative results concerning the neurofeedback approach. 14
Serious games
Rohani et al. developed a virtual classroom environment to treat children. 15 They created it with Unity3D game engine and Microsoft Kinect. Five other studies also used this platform. Ochi et al. developed a serious game with neurofeedback for adults' attention training. 16 Chen et al. also used Unity3D and NeuroSky MindWave on a serious game. 17 A significant enhancement was observed in attention and concentration during the study. Still using Unity3D, Alchalcabi et al. designed a virtual reality game with the EGG EMOTIV EPOC+. 18
The third study showed a reduction in inattention symptoms, followed by differential brain network reorganization. 19 The same research group carried out a new study and concluded that at least 20 sessions are required to start changing brain waves. 20 Machado et al. 21 developed a game called Neurofeedback Space. The results reveal signs of improvement in attention. Finally, Park et al. 22 used Unity3D to improve readability. An interactive narrative based on fairy tales and motion detection technology was used as a strategy.
In contrast, some researchers did not use neurofeedback. Dovis et al. trained several roles of children through Braingame Brian. 23 Other authors highlighted that regular computer-based cognitive training can improve some of the cognitive ADHD symptoms and still be helpful to treat video game addiction. 24
Bul et al. developed the serious game Plan-It Commander, designed to promote behavioral learning in everyday life situations that are problematic for children with ADHD. 25 Parents reported a significant improvement in time management, planning, and frustration tolerance issues. Later on, the researchers identified that boys with higher conduct disorder could benefit most from this treatment. 26
Finally, García-Baos et al. 27 used a serious game called RECOGNeyes, associated with eye tracking, to assess the impact on attention on children with ADHD. The results indicated a reduction in impulsivity.
Remote treatment
Simons et al. tried to discover the views of patients, parents, and health professionals about remote monitoring systems. 28 The company QbTech developed a prototype to monitor symptoms and side effects of the medication prescribed. Sibley et al. analyzed Cisco's Webex video conferencing tool. 29 Families reported a high level of satisfaction with the experience, and psychotherapists observed improvement for 50 percent of families. In another approach, Catalyst Common View was evaluated. 30
Other computer technologies
In wearables, Schoenfelder et al. used Fitbit Flex, a smart wristband that allows physical monitoring data. They also evaluated a social network as a form of engagement. 31 Gomez and Carro presented AdaptADHD, an app to support adaptive training and patient evaluation during their therapies. 4 The application uses a touch screen desk and aims to assist patients in improving their concentration and impulse control. Schuck et al. evaluated the usefulness of iPads and a web-based application called iSelfControl. The system was designed to support classroom behavior management. 32
Bruce et al. appraised the change in the ability to perceive danger in young drivers with ADHD who received training through Drive Smart. 33 The risk perception skills of patients improved significantly and persisted within 6 weeks of follow-up.
Web technology was also assessed. The authors sought to verify the effect of an educational site on parental perceptions. 85.5 percent of the participants considered the use relevant. As an overall result, the authors found that parents showed greater knowledge of ADHD after using the site. 3
In another approach, Biederman et al. 34 evaluated the effectiveness of an intervention based on short message service to improve the rate of adherence to stimulant drugs in adults with ADHD. The results showed that 68 percent of the intervention group replenished their prescriptions in a timely manner, against 34 percent of patients in the control group.
Mahmoodi et al. 35 developed a gamepad for computer games. Instead of hands, it can be used with feet. This difference leads to a higher level of concentration of the player.
Finally, the only use of software to train attention without success. The research investigated a system called ACTIVATE™ that targets a wide range of cognitive functions. 36
Computer technology for diagnosis
Machine learning was present in 61 percent of the studies analyzed in this SLR. Neurofeedback was ranked second with 9.3 percent participation, followed by serious games and eye tracking with 5.6 percent each. Virtual reality and actigraphy were also analyzed. This information answers the research question RQ1.
Machine learning
Magnetic resonance imaging (MRI) and its variations represent the predominant type of employed data. However, it is interesting to notice the variability found. Figure 1 shows the distribution.

Data types used in diagnosis.
In regard to the research question RQ3, some consideration is necessary. Larger samples would be expected to be related to better performance in the accuracy measured. Nevertheless, such a relationship could not be verified. Figure 2 illustrates the apparent lack of relationship between sample size and the alleged accuracy.

Sample size versus alleged accuracy of the surveys.
Recent breakthroughs in neuroimaging technology have enabled the search for biomarkers for ADHD, whether functional, connectivity, or structural as cortical thickness or brain volume. However, MRI typically generates a massive number of features. Coupled with the limited number of databases available, machine learning still faces difficulties finding the feature that would be a biomarker for ADHD. As a result, research suffers from the difficulty of interpreting models and with overfitting. Overfitting is possibly one of the reasons behind the apparent lack of correlation observed. Such a fact has also been detected by Pulini et al. 37 Other studies also highlighted the need to use more extensive databases.38–41
Given the difficulty of building a quality database, 51 percent of the researchers chose the ADHD-200. While this allows a comparison between results, in addition to providing scientific reproducibility, however, it makes the survey more limited. Table 4 shows the databases identified.
Databases Used in the Surveys
DWSMB, Dean–Woodcock Sensory Motor Battery.
Support vector machine
Zhang et al. proposed a classification framework using sparse representations. 42 Guo and He also employed support vector machine (SVM). 43 They observed differences in some regions in people with ADHD. Another article highlights such fact, suggesting that the brainstem is responsible for the accuracy of 93 percent of the individual diagnostic prediction. 44
The research conducted by Iannaccone et al. pointed out that the most predictive regions for ADHD patients were identified in the posterior, temporal, and occipital cingulate cortex. 45 SVM classifier was applied as implemented in the PROBID software. Park et al. sought to explore the differences in connectivity between ADHD subtypes. 46 Substantial differences were identified in the frontal, cingulate, parietal cortices, and the cerebellum. Likewise, Xiao and colleagues proposed a new integrated resource classification and selection framework that uses cortical thickness features from magnetic resonance data. 47
However, many believe that ADHD seems not to be related only to individual brain sections but also to the connections between them. For this reason, network-based diagnostics has attracted attention. Du et al. analyzed a discriminative subnetwork selection method to discriminative subnetworks from all brain networks. 48 The Graph Kernel Principal Component technique was applied to extract the features of these discriminative subnetworks.
Dey et al. tried to diagnose by building functional brain connectivity networks. 49 Interestingly, detection rates are higher when the classification is performed in male and female groups separately. In another investigation, raw features were derived from temporal variability among the intrinsic connectivity networks. 50
The PROBID software was also employed by Hart et al. to predict diagnosis based on task-based activation patterns. 51 Still evaluating the use of the SVM classifier, Chaim-Avancini et al. sought a neuroanatomical signature associated with the ADHD spectrum in adults. 39 Dea et al. 52 collected electroencephalographic data during the sleep period of 18 children, followed by linear SVM for classification.
Variations in the use of SVM classifier
Shao et al. proposed a bi-objective ADHD classification scheme based on SVM classifier. 41 Instead of the traditional weighted sum formulation, a bi-objective SVM formulation was adopted. Miao and Zhang discussed the ADHD classification using the preprocessing and fractional amplitude of low-frequency fluctuation in resting-state fMRI (rs-fMRI) data. 53 The Relief algorithm seems to be one of the best for feature selection, as it has high efficiency and more satisfactory results. Riaz et al. evaluated the combination of image and non-imagery data. 54 As a general result, data fusion demonstrated improvement in diagnostic performance.
Finally, Joy et al. 55 used the EEG data for diagnosing ADHD, together with the SVM classifier. A substantial difference between the expected and abnormal EEG signals was observed in the minimum values of the spectral bands α, β, γ, δ, and θ.
Other approaches
Qureshi et al. used the hierarchical extreme learning machine (H-ELM). 56 They observed that a recursive feature elimination approach combined with H-ELM allowed the acquisition of high multiclass classification accuracy rates. Also, the most important features were the upper frontal lobe surface area, cortical thickness, volume, and mean surface area of the entire cortex. In 2017, the authors calculated the global mean functional connectivity maps across the cortex to extract anatomical features. Cortical surface reconstruction and volumetric segmentation were performed with FreeSurfer. 57
He et al. developed a Multi-way Multi-level Kernel model that can extract discriminative, nonlinear, and structural preserving representations of tensor data. 58 By using data from the ADHD-200 competition, they achieved diagnostic results of 70 percent accuracy.
Some articles have highlighted the sensorimotor deficits present in ADHD patients. Finch et al. reported using the Dean–Woodcock Sensory Motor Battery (DWSMB) method. The DWSMB uses many classic sensory and motor function confrontational tasks. 59
Itani et al. explored the use of decision tree based on a multilevel approach. 60 That is, the focus is on diagnosing ADHD individuals and then distinguishing among the three existing ADHD subtypes. The implementation of the C4.5 algorithm proposed by Weka software was considered.
Still exploring the use of decision trees, another research used these algorithms. The data set used consists of eye movements and positions, in addition to the different types/events of the look. The product used for eye tracking was the Tobii Pro X2-60. 61
Leontyev et al. 62 employed a new version of the Stop-Signal Task with machine learning techniques combined with mouse tracking. Moreover, Lanka et al. 63 applied 18 different machine learning classifiers. The results reinforced the understanding that overfitting is a massive problem with heterogeneous and small-sized databases.
Naive Bayes
Cicek et al. evaluated K-nearest neighbors (K-NN) algorithm and naive Bayes using Matlab. 64 Lg et al. also analyzed naive Bayes and J48 classifiers in ADHD diagnosis. 65 The results showed that the J48 algorithm had better results than naive Bayes, but it demanded more time for execution. Eslami and Saeed also applied the K-NN through a model selection scheme designed to pick the optimum value of k from the training data. 66
Use of neural networks
Deshpande et al. appraised fully connected cascade architecture. 67 The most discriminative connectivity features showed reduced and altered connectivity involving the left orbitofrontal cortex and various cerebellar regions in ADHD. Other research analyzed three-dimensional (3D) convolutional neural networks in rs-fMRI data. The approach obtained superior performance when compared with the two-dimensional classifiers. 68
Rahadian et al. compared the use of genetic algorithms combined with neural networks to aid initial calibration of the weights. 69 The technique is known as learning vector quantization 2 neural network. The results pointed to a 10 percent higher accuracy than the pure use of learning vector quantization. Another group of researchers employed deep belief networks using more than fMRI images, such as age, IQ, and functional features. 70
Finally, the hippocampus region was used as a projection-based learning algorithm for a meta-cognitive radial basis function network. The alleged accuracy suggests the use of the hippocampus as a viable structure for diagnosis. 71
Other types of data
Considering that ADHD is a highly genetic disorder, Wang et al. aimed at potential biomarkers of microRNAs (miRNAs). 72 The proposal was to identify the expression profiles of these miRNAs with the polymerase chain reaction via real-time quantitative reverse transcription. Finally, 13 miRNAs were identified as potential ADHD biomarkers.
Uluyagmur-Ozturk et al. investigated the diagnosis based on the recognition of emotions through facial expressions. 73 Other research proposed the use of time–frequency features of neuroelectric activity related to events. The time–frequency Hermite atomizer technique was used for the collection of high-resolution time–frequency features. 74
Anderson et al. worked with a multimodal neuroimaging framework. A subset of features from each information source was mapped to one or more dimensions and interpreted using generative models. 75 Structural measurements of the default mode network (DMN) regions were related to diagnosing ADHD inattentive type. Ventral DMN subnetworks may have more functional connections in ADHD-I, whereas dorsal DMN may have less. In another approach, Yasumura et al. used the near-infrared spectroscopy imaging technique to calculate the change in oxygenated hemoglobin in the prefrontal cortex. 76
The SLR also verified the use of neuropsychological measures (interviews and questionnaires). A specific study evaluated machine learning in predicting ADHD. The authors claim 100 percent accuracy. However, the sample size used was very low. 47
Finally, Crippa et al. investigated the ability of multiple domain measurement profiles, including blood fatty acids, neuropsychological measurements, and functional near-infrared spectroscopy measurements. 77 The analysis of each data set separately suggested that the most discriminative blood fatty acids were the linoleic acid, and the total amount of polyunsaturated fatty acids.
Table 5 summarizes the machine learning techniques.
Machine Learning Technologies
DBN, deep belief network; DTI, diffusion tensor imaging; FCC, fully connected cascade; fMRI, functional magnetic resonance imaging; fNIRS, functional near-infrared spectroscopy; GA-LVQ2NN, genetic algorithm learning vector quantization 2 neural network; H-ELM, hierarchical extreme learning machine; LASSO, least absolute shrinkage and selection operator; miRNA, microRNA; MMK, Multi-way Multi-level Kernel; MRI, magnetic resonance imaging; NA, not applicable; NIRS, near-infrared spectroscopy; NMF, non-negative matrix factorization; PBL-McRBFN, projection-based learning algorithm for a meta-cognitive radial basis function network; RFE, recursive feature elimination; R-RELIEF, reliable RELIEF algorithm; rs-fMRI, resting-state functional magnetic resonance imaging; VA, verification accuracy.
Other computer technologies
Faraone et al. explored a game called Groundskeeper to perform diagnosis. 78 Another study developed a system that assesses symptoms through multidimensional data, including physiological, motion, and task-related data. 79 Mwamba et al. 80 experimented with a game developed for the tablet through the Unreal Engine platform. The results were submitted to the SVM classifier.
Neurofeedback was analyzed by Lee et al. over a combination of virtual reality, eye tracking, and EEG technologies. 81 Selective attention, sustained attention, abstract reasoning, and cognitive transference skills were assessed in a 3D virtual classroom with distraction factors. The HTC Vive, which is a virtual reality device, was used. The aGlass DKII was selected for eye tracking. Finally, the MindWave was used as an EEG device. The same research group seems to have evolved the research work by making some adaptations in developing a virtual classroom environment. 82
The same EEG device was used by Garcia et al. 83 to identify the level of attention based on alpha and beta brain waves. There was also the implementation of graphical visualization of brain waves using the LABVIEW software. Another study also explored neurofeedback, this time using the portable data monitor EEG EMOTIV EPOC+, associated with Arduino UNO to generate auditory stimuli, with the subsequent capture of the data by the device. 84
Gilbert et al. examined the diagnostic accuracy of actigraphy to measure the non-task-related motion. 85 To do so, two motion sensor devices ActiGraph GT3X were used on the individuals (one on the wrist and the other on the ankle).
Other studies examined information systems for diagnosis. Kyeong et al. investigated applying a topological data analysis tool called Mapper to analyze the functional brain connectivity data. 86 Bart et al. evaluated the reliability of an online continuous performance test through a Web system. 87
Two recent surveys have made use of specific applications. Hernaiz-Guijarro et al. 88 used the DMDX software to provide a methodology for analyzing children's responses about tasks with visual stimuli. While Unal et al. 89 performed three computerized tests for the diagnosis: Stroop Test, Stroop Plus, and Perceptual Selectivity. The Stroop Test provided the best descriptive measure.
Table 6 summarizes the technologies used in the ADHD diagnosis.
Technologies Used in the Diagnosis
fALFF, fractional amplitude of low-frequency fluctuation; GPC, Gaussian process classifiers; K-NN, K-nearest neighbor; MALINI, machine learning in neuroimaging; OCPT, online continuous performance test; RF, random forest; RFE-SVM, recursive feature elimination-SVM.
Conclusions
This study listed the leading computer technologies being researched in both ADHD diagnostic and treatment phases, as well as their limitations and trends. It was also possible to observe a wide variety of data types explored as input to the algorithms, demonstrating the continued search for a biomarker. Another contribution of this SLR is about data analysis currently used in machine learning research and its restrictions. For instance, the research emphasized the apparent lack of relationship between sample size and the alleged accuracies, which certainly contradicts the expectations for machine learning. Finally, this study highlighted the convergence of evidence from different studies on the persistence of long-term effects of neurofeedback in treating ADHD.
Footnotes
Authors' Contributions
R.M.B.A. and M.F.d.S. conceived and designed the analysis, and wrote the article. E.A.S. and A.J.A. collected the data and revised English.
Author Disclosure Statement
No competing financial interests exist.
Funding Information
No funding was received for this article.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
