This study analyzed research that used ML models to predict ADHD in children. The variables were: ADHD diagnosis (target) and EEG-derived data (predictors). The guidelines for the systematic review and meta-analysis were developed and managed with CADIMA Centro Leibniz de Investigación del Paisaje agrícula [ZALF, 2024]. The methodology followed PRISMA guidelines (Page et al., 2021). The WoS and Scopus repositories were used to track publications, with a viewing window from January 2012 to July 2024. Research applying ML techniques to the analysis of EEG signals for the diagnosis of ADHD began to be published in 2012. There were no language or subject area restrictions in the study. Some of the disciplines included were medicine, neuroscience, computer science, engineering, psychology, genetics, social sciences, biological engineering and molecular biology. The Medical Subject Headings (MeSH) (National Library of Medicine, 1960) were used to identify appropriate search terms. The exploration was performed by an expert group in neuroscience and engineering. The equation used in Scopus (TITLE-ABS-KEY) and WoS (All Fields) is presented below.

To minimize bias in the publications included in the meta-analysis and SLR, precise and rigorous inclusion criteria were established: (1) empirical and structured studies, (2) focused on the diagnosis of ADHD with ML in children, (3) using EEG data as predictor variables, (4) employing training and validation samples, (5) using ML-specific metrics to assess model performance, (6) open access publications. Exclusion criteria were: (1) studies in adults, (2) without data standardization, (3) literature reviews, single cases, memoirs, monographs or technical manuals, (4) studies with data from psychometric testing, spectrograms, genetic data, MRIs actigraphy, accelerometer or virtual reality, (5) no age reporting of sample groups, (6) ADHD with comorbidities or associated with genetic syndromes, (7) unsupervised learning ML models, (8) with Accuracy less than 65%.

Two reviewers independently assessed the title and abstract according to the established eligibility criteria. Any disagreement was discussed with the third reviewer and resolved by consensus. An adapted checklist designed to ensure that the selected publications were relevant to the object of study of this SLR was used (Mohamed Shaffril et al., 2021). The checklist included: (a) study design, with a description of the problem statement and the inclusion requirements of the research units; (b) construction of the ML predictive models that contemplated the same target variable; (c) analysis of results, focusing on model performance, description of statistics, and conclusions; and (d) interpretation with generalization of findings, limitations, and formulation of new hypotheses.

Initially, 28 publications were identified in Scopus and 39 in WoS, totaling 67; 15 duplicates were eliminated from the two search engines, leaving 52 publications in total. The three reviewers considered 30 studies that met the eligibility criteria adequate and discarded 22 for the following reasons: (1) studies in adults: 6, (2) no data standardization: 0, (3) literature reviews, single cases, memoirs, monographs, meta-analyses or technical manuals: 5, (4) studies with psychometric test data, spectrograms, genetic data, MRI, actigraphy, accelerometer or virtual reality: 2, (5) ADHD with comorbidities or associated with genetic syndromes: 4, (6) without reporting age of sample groups: 1, (7) unsupervised learning ML models: 3, and (8) with precision (accuracy) less than 0. 65: 1. Twenty-six articles were included in the meta-analysis that reported the complete metrics of the optimal ML model: Accuracy, Sensitivity (Detection) and Specificity (Exclusion). Four articles were eliminated that, although reporting accuracy, did not include sensitivity (2) or specificity (4). Figure 1 presents the PRISMA flowchart (Page et al., 2021).

The sample groups, age of participants, ML models, performance metrics, validation, EEG waveforms and evaluated condition (cognitive task, closed eye and open eye) of the 30 selected publications are presented in Table 1. The order of tabulation of the studies was assigned according to the ML model constructed: SVM models appear first (9), followed by CNNs (4), ANNs (2), k-NNs (2) and Naïve Bayes (2). In Table 1, starting from row 20, 11 articles are reported that employ different techniques: K2 Bayesian networks (1), ConvMixer-ECA (1), eXtreme Gradient Boosting (XGB) (1), Explainable Boosting Machine (EBM) (1), Gaussian Process with Radial Basis Kernel (GF-RBF) (1), Generalized Linear Model (GLM) (1), Sparse Non-Negative Least-Square coding (SNNLSC) (1), Monte Carlo Simulation (1), Bagged Tree (BT) (1), CatBoost (1) and Random Forest (RF) (1).

From the selected research, it was identified that the applying of ML for the diagnosis of ADHD with EEG has increased in recent years. Although the first papers date back to 2012, it was not until 2023 that this trend began to consolidate. In 2017 there was one publication, and in 2018 two. Interest increased in 2019 with three publications, and in 2020 it remained at two. In the following years, there were fluctuations, with ten publications, marking the importance of strengthening the use of ML for this diagnosis. In 2024, there were five publications, reflecting a sustained interest in this area of research (Peterson et al., 2024).

Of the 30 publications chosen, sample groups ranged from 10 to 135 participants. Appropriate sample size is critical for training ML models; external validation studies with insufficient sample sizes may yield inaccurate estimates of model efficacy, affecting calibration, discrimination and clinical utility. Studies should include at least 100 cases to ensure accurate validation (Riley et al., 2021). The diagnosis of ADHD requires identifying the three clinical presentations (inattentive, hyperactive-impulsive or combined). Only four of the papers reported ADHD presentation in their analyses (Bashiri et al., 2017; Chen, Chen, et al., 2019; Li et al., 2023) and 28 studies included clinical groups (ADHD) and healthy controls. The inclusion of a control group allows reliable comparisons between minors with ADHD and those with normal development and provides a baseline to assess differences in cognitive and neurophysiological metrics (Zhao et al., 2020). Regarding sex information, 16 studies mentioned the distribution between girls and boys (Alkahtani et al., 2023; Feng & Xu, 2024; Karabiber Cura et al., 2023).

Previous studies have highlighted that minor girls with ADHD are often diagnosed later than minors boys, due to the prevalence of emotional clinical manifestations that often hinder the identification of ADHD symptoms. Referring to gender is important, because of differences in clinical presentation and symptomatology that influence diagnosis and appropriate treatment (Klefsjö et al., 2020; Mayes et al., 2020). The modal age range in the selected publications was 7 to 12 years, which coincides with studies indicating that it is in this period when ADHD symptoms become more evident, cease to be masked and begin to generate difficulties in the family and school context (Espinet et al., 2022; Rocco et al., 2021).

SVMs were employed in nine studies to accurately classify ADHD diagnoses, process large volumes of data and create decision boundaries in high-dimensional spaces (Pisner & Schnyer, 2020). This technique has shown high levels of sensitivity and specificity in differentiating between minors with ADHD and those with normal neurodevelopment, which can address the heterogeneity of ADHD symptoms. However, SVMs are at risk of overfitting when the number of features is high compared to sample sizes, making interpretation difficult (Bledsoe et al., 2020). In four studies employing CNNs, they showed an average accuracy (Accuracy) of 96.24%, outperforming other ML models, such as ANNs and k-NNs (Mafi & Radfar, 2022; Moghaddari et al., 2020; TaghiBeyglou et al., 2022). Unlike SVMs, which require manual feature selection and may be limited by data linearity, CNNs can capture more complex relationships by processing EEG data into time-frequency representations without prior feature engineering; this reduces the risk of bias and optimizes diagnostic accuracy (Chen, Song et al., 2019).

Cross-validation (k-fold) is one of the methods to assess the performance of ML models (Ramezan et al., 2019). Validation with 10 layers (k = 10) was employed in thirteen of the selected studies, i.e., the data were divided into 10 subsets. They were trained with nine of them and validated with the remaining ones, reiterating this process 10 times to ensure that each subset is taken up for validation at least once (Alim & Imtiaz, 2023; Attallah, 2024; Muthuraman et al., 2019). This method allows the accuracy of the model to be evaluated robustly and avoids problems with overfitting, diagnostic reliability is increased (Chen, Chen, et al., 2019; Moghaddari et al., 2020). In eleven investigations, a validation with 5 layers (k = 5) was used, where the data is classified into 5 fragments, although this approach reduces the number of partitions, it allows a proper evaluation of the model performance, with lower granularity compared to that of 10 layers (Abibullaev & An, 2012; Alsharif et al., 2024; Li et al., 2023). The purpose of the layers is to reduce bias and variance in the model results and ensure that new data are generalized (Lei, 2019).

In 22 of the selected studies, different wave frequencies (δ, θ, α, β, γ) were used to evaluate eyes closed (EC), eyes open (EO), and cognitive task conditions in experimental groups (ADHD) to analyze the neurophysiological responses under each condition (Ahire et al., 2023; Alkahtani et al., 2023; Maniruzzaman et al., 2023). EEG is a noninvasive monitoring method that detects brain electrical impulses through electrodes placed on the scalp (Abdulwahab et al., 2020). During the OC condition the infants in the experimental group (ADHD), presented an increase in theta (θ) waves and a decrease in beta (β) waves, which contrasts with the typical pattern of brain relaxation and synchronization observed in typically developing infants (Alkahtani et al., 2023).

In the EO condition, it was found that in ADHD there are elevated levels of hyper-coherence in beta (β) frequencies, suggesting excessive activity in areas responsible for visual attention and external information processing, reflecting rigidity in response to visual stimuli (Deshmukh et al., 2024; Kerson et al., 2023). In the cognitive task condition, selected investigations used tests of visual attention and found that infants with ADHD present pronounced activation in the alpha wave (α = 8-12 Hz), suggesting increased distractibility and extra effort for attentional sustainment. Increased theta wave (θ) power was evident, indicating difficulty in maintaining concentration when confronted with visual stimuli. The elevated theta-to-beta (θ/β) ratio in these children shows an altered neural balance, hindering the ability to focus and process information (Abibullaev & An, 2012; Öztoprak et al., 2017; TaghiBeyglou et al., 2022).

To perform the meta-analysis, the accuracy, specificity and sensitivity metrics of the selected papers were used. The analysis was developed in Python (Python Software Foundation, 2023, version 3.12’) (Appendix). First, the data from the 26 studies that recorded the three metrics were entered into the software. Second, the values of the descriptive statistics for each measure were obtained, see Table 2.

Third, a “Forest Plot” diagram was developed to visualize the confidence intervals of the averaged metrics (accuracy, sensitivity, specificity) from the studies and to analyze the variability (Figure 2).

An interval plot is a graphical representation commonly used in meta-analysis to visualize the findings of multiple investigations, the black dot on the line indicating the value of the metric analyzed. Longer horizontal lines indicate greater uncertainty in the estimate of effect (specificity) (Dettori et al., 2021). Most of the articles have an accuracy > 94%, a sensitivity > 94% and a specificity > 92%, which allows us to conclude that the studies analyzed perform well in terms of accuracy, sensitivity and specificity, see Figure 2.

Table 3 presents the results of the pooled effect and heterogeneity. The pooled effect represents the weighted average of the individual studies (Paul & Barari, 2022). The pooled values for accuracy, sensitivity, and specificity were 98%, 95%, and 96%, respectively, indicating good performance.

To measure the heterogeneity of the studies, we used the descriptive Q (Equation 1). A value of Q = 0.00 means no variation among the studies, i.e., they are completely homogeneous. The average effect was obtained from the mean weighted by the inverse of the variance of the effect of each study (Equation 1) (Huedo-Medina et al., 2006).

Where:

− w i = 1 Var ( θ i ) (the inverse of the variance of the i-th study’s effect estimate).

− θ ^ i is the effect estimate of the i-th study.

-θ is the overall pooled effect estimate, calculated by (Equation 2):

To calculate the percentage (%) of variability between studies attributed to heterogeneity, the I² statistic was used (Equation 3). This indicates the percentage of total variability due to between-study heterogeneity (Stogiannis et al., 2024). An I² of 0.00% suggests no heterogeneity, i.e., the studies are consistent.

Where:

- Q is the heterogeneity statistic previously calculated.

- k is the total number of studies included in the meta-analysis.

- An I² of 0.00% suggests that there is no heterogeneity, meaning that the studies are consistent

The p-value for the Q statistic is obtained from the chi-square statistic, with 25 degrees of freedom (n-1). A p-value of 1 indicates that there is no evidence of significant heterogeneity between studies. The absence of heterogeneity (Q = 0.00, I² = 0.00%) shows that the individual studies analyzed are highly consistent, which is generally positive because it suggests that the results are reliable and generally applicable.

The heterogeneity analysis revealed that the studies included in the meta-analysis were highly consistent, with Q = 0.00 and I² = 0.00%, indicating no significant variability among the estimated effects. A fixed-effect statistical model was used for the pooled result, all studies shared the same true effect and the differences observed were due exclusively to sampling error. The fixed-effects model in meta-analysis assumes that all studies share the same true value of the estimated parameter, so any observed variability is due solely to sampling error (Barili et al., 2018; Borenstein et al., 2010). This approach assigns greater weight to studies with lower variance and is appropriate when heterogeneity is minimal (I² close to 0%), such as in this work

This decision implies that the results obtained are generalizable only to the set of studies analyzed and to populations with similar methodological characteristics. In contrast, a random-effects model, which considers that the studies represent a sample drawn from a larger population with variability among them, was not necessary because of the homogeneity observed. From a conceptual perspective, the homogeneity among the included studies is justified by the use of comparable methodologies in predicting ADHD with EEG data and machine learning models, employing consistent metrics such as accuracy, sensitivity, and specificity. These factors reinforce the robustness of the results and their validity within the frame of reference defined by the meta-analysis. The findings of the meta-analysis indicate that ML techniques applied to the diagnosis of ADHD using EEG are effective and reliable as complementary tools in the clinical setting (Appendix).

ML techniques are effective in classifying ADHD in infants by analyzing EEG data. The analyzed models achieved significant accuracy rates (94.92%), indicating that they can capture distinctive neural patterns associated with this clinical condition. This favors diagnostic accuracy compared to traditional clinical methods.

The papers included in the SLR reported that theta (θ) waves and beta (β) waves are predominant in EEG analysis for the detection of ADHD. In this population, increased theta (θ) wave activation has been observed, associated with distractibility and inattention. Beta (β) waves, related to cognitive activity and focus, show a reduction, which reflects a deficit to control responses and maintain concentration for a prolonged period.

The main limitations of the analyzed publications include the small size of the different sample groups, which could compromise the extrapolation of the findings to the global population. Most of the articles do not classify ADHD according to its presentations (inattentive, impulsive/hyperactive or combined), which affects diagnostic accuracy. Another limitation is that not all research reports sex (boys/girls) when making classifications, which biases the results. The findings have not been tested in practical settings, which reduces direct applicability to the clinical setting. Future work could focus on the integration of multimodal data such as EEG, MRI, behavioral and neuropsychological data, to generate an accurate and precise diagnosis of ADHD.