Computer-aided diagnostic tools based on encephalography and deep learning for the detection of mental issues: Case of attention deficit hyperactivity disorder

• Autores: Javier Sanchis Bernabéu
• Directores de la Tesis: Juan Trujillo Mondéjar (dir. tes.) , Miguel A. Teruel Martínez (dir. tes.)
• Lectura: En la Universitat d'Alacant / Universidad de Alicante ( España ) en 2024
• Idioma: inglés
• Número de páginas: 196
• Tribunal Calificador de la Tesis: Antonio Fernández Caballero (presid.) , Manuel Marco Such (secret.) , Maribel Yasmina Campos Alves Santos (voc.)
• Enlaces
  • Tesis en acceso abierto en: RUA
• Resumen

  • Attention-Deficit/Hyperactivity Disorder (ADHD) is a prevalent mental health issue that significantly affects daily functioning and quality of life, especially in children and adolescents. Traditional methods of diagnosing ADHD are based on the DSM-5 criteria and tend to be subjective, relying heavily on symptoms reported by patients or their parents. In addition, these methods are supported by tests that can give inaccurate answers, affecting the accuracy of the diagnosis. The process can also be lengthy, as the DSM-5 requires at least 6 months of symptoms for diagnosis confirmation. These challenges highlight the need to implement more objective and less time-consuming diagnostic tools. The development of neural networks for ADHD diagnosis using biometric signals, such as electroencephalography (EEG), can help clinicians address these challenges. By employing a signal that directly reflects brain activity, it is possible to provide more objective diagnoses based on measurable data. Once the EEG signal is obtained, the computer system can generate results in a matter of seconds, as the neural networks have been pre-trained. Moreover, these networks can be continuously updated with new data from diagnosed subjects, thus improving the accuracy of the diagnosis. Finally, such systems could also be widely implemented, for example, in schools, allowing a larger number of potential ADHD cases to be detected for further evaluation by a specialist. It should be noted that these systems are not intended to replace physicians, but rather to help them detect more cases, thus enabling them to treat a larger number of people. Although several works have been developed in the field of neural networks and EEG, the results can be improved and the tools can be made more reliable and replicable. Another technology that could help address the inherent limitations of traditional diagnostic methods is serious games. These gamified systems can capture real-time data on the responses and decision-making abilities of players, providing a more objective assessment of symptoms such as inattention and impulsivity. In addition, video games may not be perceived as a form of treatment or an imposition by parents or teachers, potentially making them more appealing to children. Furthermore, according to the DSM-5, no specific cause for ADHD has been identified, which has generated interest in studying the underlying neural mechanisms. However, studies on the brain regions associated with ADHD show heterogeneous results, pointing to different hemispheres and lobes. This divergence underlines the complexity of ADHD and highlights the need for further research. Finally, with a better understanding of the origin of ADHD, more cost-effective and specialized EEG signal capture devices could be developed. To address the aforementioned challenges, the main goal of this doctoral thesis is to explore the development and implementation of computer-aided systems, leveraging computer engineering and artificial intelligence, to support the diagnosis and enhance the understanding of ADHD. From this objective, three lines of research are defined: (i) developing neural networks using EEG signals to assist in the diagnosis of ADHD, (ii) identifying the brain regions that best characterize ADHD, and (iii) creating serious games as diagnostic tools and biometric data collection systems to detect ADHD in an objective and reliable manner. First, a neural network was developed that achieved an accuracy of 88.42%, outperforming existing models. The hyperparameter tuning of this network also reduced the training time from 15.5 days to 1.3 days, without compromising performance. This tuning enabled the development of new architectures by focusing on key elements of the neural network, leading to a second, more accurate network for diagnosis, achieving 94.87% accuracy. Secondly, this work identified key brain regions relevant to ADHD diagnosis by analyzing various subsets of EEG channels. The results showed that the Frontal Lobe and Left Hemisphere provided significant information for classifying ADHD, achieving f1-scores of 80.81% and 80.56%, respectively. Feature selection methods, such as Backward Stepwise Selection, improved classification performance to an f1-score of 82.81%. Finally, the Serious Game “Attention Slackline” was designed to contribute to the assessment of attention with objective and systematic measures. During the game, intrinsic metrics such as hits, omissions, and commissions were collected. In addition, biometric data such as EEG, electrodermal activity, and eye tracking were recorded, providing a high-quality dataset for future research. To test the feasibility of the game as a tool for measuring attention, a protocol was defined in collaboration with a team of psychologists. The aim was to test whether the results obtained from the metrics obtained during the game correlate with those obtained from the well-known D2 attention test. As we obtained positive correlations, it is confirmed that “Attention Slackline” is a valid tool for measuring attention. The findings of this thesis indicate that integrating AI techniques and serious games into ADHD diagnostics can lead to more accurate, efficient, and engaging assessments. Future work will focus on exploring other DL architectures, incorporating multimodal data, developing personalized models, and ensuring the explainability of these AI systems to facilitate their integration into clinical practice.