Attention-Deficit Hyperactivity Disorder (ADHD) ranks among the most prevalent mental issues diagnosed in childhood

The gap

Evidence profile

Sourced from the future work of the source papers, classified as general, spanning 3 journals.

Supporting evidence — 3 representative gaps

• An adaptive epilepsy detection system using federated learning (2026) · Multimedia Tools and Applications · doi

  This study developed and evaluated an Adaptive Multi-modal FL system for the detection of epilepsy using EEG and MRI data. By utilizing specialized CNN architectures, the pro- posed system successfully addressed the challenges of data privacy and heterogeneity across 14 simulated clinical clients. The experimental results demonstrated that the FL approach not only exceeded the diagnostic accuracy of centralized benchmarks achieving 98.53% for EEG and 91.35% for MRI but also significantly reduced training latency by up to 88%. Ulti- mately, this research validated that decentralized learning could provide a robust, privacy- compliant, and computationally efficient solution for real-world clinical seizure detection. Furthermore, the system delivered substantial operational efficiencies, notably in training time: EEG training was reduced from 7 hours, 43 minutes, and 3 seconds to 3 hours, 24 min- utes, and 8 seconds, and MRI training was reduced from 5 hours, 9 minutes, and 2 seconds to just 59 minutes and 6 seconds. This significant time reduction underscores the scalability and computational viability of the FL approach for real-world clinical deployment. These findings underscore the potential of FL to advance seizure detection technologies by con- currently addressing two paramount concerns in medical artificial intelligence: robust data privacy and operational efficiency. The proposed system offers a promising solution for inte- grating advanced Deep Learning techniques into routine clinical practice, paving the way for future research focused on personalization and broader applicability in healthcare settings. Building on the findings of this study, several actionable directions for future research are identified. Future work will extend the proposed system by exploring PFL to tailor mod- els to individual patient seizure patterns, further enhancing detection accuracy for diverse clinical profiles. Integrating additional data modalities, such as wearable sensor data and Magnetoencephalography (MEG), will support more comprehensive and continuous patient monitoring. Furthermore, we will focus on optimizing FL algorithms to handle extreme data heterogeneity across larger client networks, utilizing techniques such as FedProx aggrega- tion, model compression, and asynchronous updates to reduce communication overhead and enhance scalability. Finally, incorporating advanced security layers like Differential Privacy (DP) and conducting real-world clinical validation across diverse hospital environments will be crucial to ensure regulatory compliance and guarantee the system’s practical impact on automated seizure detection technologies. Acknowledgments The authors would like to express their sincere gratitude to Damietta University for its generous financial support in covering the publication fees of this research. This support has been invalu- able in facilitating the dissemination of our work, and we deeply appreciate the University’s commitment to promoting scientific research and academic excellence. Author Contributions The authors equally contributed to this work, and all have reviewed and approved the final version of the manuscript.

  generalfuture work
  Keywords: system clinical detection privacy training seizure seconds across reduced real world hours minutes future support
• Multiscale deep learning convolutional neural network for ADHD detection using EEG (2026) · Multidimensional Systems and Signal Processing · doi

  Attention-Deficit Hyperactivity Disorder (ADHD) ranks among the most prevalent mental issues diagnosed in childhood. In recent years, several computer-aided systems for diagnosing ADHD using EEG and Deep Learning have been developed. How- ever, the current state of the art presents three main drawbacks. The majority use manual feature selection. They manually remove the artefacts from the EEG signals, preventing the user from performing an automatic end-to-end classification process. Finally, they use an inter-subject validation approach, which is unsuitable in this case (over-fitting problem). To solve the above-mentioned issues, we proposed an efficient Deep Learning Multiscale Convolutional Neural Network for detecting ADHD using EEG data (EEG- MSCNet). The proposed approach performs automatic feature extraction from the EEG signals. EEG-MSCNet is evaluated by performing a Leave-One-Out subject- cross-validation on the EEG dataset; the subjects for training and testing the model were different. The experimental results showed that EEG-MSCNet is robust and pro- duces reliable results. Consequently, it can be applied on an outpatient basis, allowing physicians to follow their patients more closely and speed up their treatment time. Furthermore, the proposed EEG-MSCNet model should be tested on other EEG datasets related to ADHD or other mental health issues to enhance its reliability. It is also essential to consider the contribution of each channel in the EEG signal to the outcome. Identifying a subset of channels that yield similar results to those obtained from using all channels would lead to reduced computation time. Moreover, identifying the channels that have the most significant impact on the outcome would provide insight into the brain regions that play a critical role in developing ADHD. This information could provide clinicians with invaluable insights into the origins of ADHD and, consequently, a better understanding of its nature. Ultimately, it is recommended to consider alternative artefact removal method- ologies. This strategic approach aims to glean additional insights into the model’s performance while enhancing its robustness. With these future works, the premature diagnosis of ADHD will be improved and enhance the quality of life of people affected by AHDH. Finally, the source code, the original database, the pre-processed dataset, and all file logs are fully available. Therefore, the results obtained can be easily validated and replicated. In addition, they can also serve as a scaffold for further research.

  generalfuture work
  Keywords: adhd mscnet issues using approach proposed model channels mental deep learning feature signals performing automatic
• Frequency-aware transformer networks for robust and generalizable EEG-based seizure detection (2026) · Scientific Reports · doi

  This study introduced BrainXNet, a multi-scale spectro-temporal attention framework for robust EEG-based seizure detection. By integrating multi-scale convolutional feature extraction, a spectral attention mechanism, and a temporal transformer encoder, BrainXNet was able to capture fine-grained local patterns, dynamically emphasize clinically relevant frequency bands, and model long-range temporal dependencies in a unified architecture. Extensive experiments conducted on the CHB-MIT and TUH-SZ benchmark datasets demonstrated that the proposed model consistently outperformed state-of-the-art baselines, achieving accuracies of 99.1% and 98.4%, respectively. In addition to excelling in clean conditions, BrainXNet proved highly resilient to noise and artifacts and maintained strong cross-dataset generalization, both of which are crucial for real-world clinical deployment. Quantitatively, BrainXNet achieved an accuracy of 99.1%, sensitivity of 98.8%, specificity of 99.4%, F1- score of 98.8%, and an AUC of 0.997 on the CHB-MIT dataset. On the TUH Seizure Corpus, the proposed framework obtained 98.4% accuracy, 97.9% sensitivity, 98.8% specificity, 97.7% F1-score, and an AUC of 0.995. Furthermore, BrainXNet maintained more than 94% accuracy during cross-dataset evaluation and demonstrated high robustness under noisy conditions, outperforming conventional CNN-, RNN-, and transformer-based approaches. These quantitative results confirm the effectiveness of the proposed multi-scale spectro-temporal attention framework for reliable and generalizable EEG-based seizure detection. The results suggest several important implications. First, the fusion of spectral and temporal attention mechanisms is a powerful strategy for modeling nonstationary biomedical signals such as EEG, where seizure activity may manifest differently across patients and conditions. Second, the robustness of BrainXNet in noisy and cross-dataset settings supports its potential for reliable application in diverse clinical environments, bridging the gap between controlled research scenarios and practical diagnostic use. Finally, the modular design not only enhances performance but also opens avenues for interpretability, as the learned attention weights could be examined to highlight clinically meaningful frequency bands and time windows. Despite these advances, certain limitations remain. The current study primarily focused on epoch-level detection, whereas patient-level or event-level evaluation would better reflect clinical workflows. Moreover, although BrainXNet was validated on two large datasets, additional testing across broader populations and acquisition setups is required before clinical adoption. Computational efficiency was shown to be competitive, but deployment in portable or wearable EEG systems may require further optimization for low-power environments. Future work will explore several directions. One avenue is the integration of domain adaptation and transfer learning techniques to further enhance generalization across datasets, institutions, and hardware platforms. Another is the incorporation of explainability frameworks, enabling clinicians to visualize which spectral– temporal regions drive model decisions, thereby fostering trust in AI-assisted diagnosis. Extending BrainXNet to real-time seizure prediction rather than detection is also a promising research direction with high clinical impact. Finally, prospective validation through clinical trials will be essential to assess the model’s effectiveness in real-world patient monitoring and decision-support systems. In conclusion, BrainXNet demonstrates that combining multi-scale convolution, spectral attention, and transformer-based temporal modeling establishes a new benchmark for automated EEG seizure detection. By addressing accuracy, robustness, generalization, and efficiency, this work provides a solid foundation for future efforts aimed at developing clinically deployable seizure detection and prediction systems.

  generalfuture work
  Keywords: brainxnet temporal seizure attention detection clinical multi scale based spectral model dataset accuracy framework transformer

Questions about this gap