Attention deficit hyperactivity disorder prediction using EEG signals and deep learning models

Attention deficit hyperactivity disorder (ADHD) is a common, comorbid neurodevelopmental disorder that begins in childhood. Diagnosis is made by professionals using behavioral rating scales, without biomarkers. Delays in accurate diagnosis and treatment result in worsening of other psychiatric conditions that co-occur with the disease and, as a result, increased utilization of healthcare services. Therefore, there is a need for potential biomarkers that can be used in the diagnosis of ADHD. In this thesis, it is aimed to propose a decision support algorithm which will reduce the dependence on subjective tests and provide rapid and successful diagnosis by examining different convolutional neural networks (CNN), one of the deep learning methods, that have high accuracy for ADHD diagnosis by using frequency-time features in EEG signals of children diagnosed with ADHD and healthy controls, and to discover how hidden information in the time-frequency domain have an impact on the diagnosis of ADHD by using explainable artificial intelligence (AI) methods. For this purpose, the data was pre-processed and noise was removed, and spectrogram images providing time-frequency information were created. Within the scope of the thesis, transfer learning was studied on CNN architectures AlexNet, GoogleNet, ResNet-50, and SqueezeNet. Firstly, CNN architectures were used with their own classifiers. The most successful performance was achieved with the AlexNet architecture with 97.78%. The signals divided into channels and frequency bands and the distinctiveness of brain regions and frequencies in diagnosis were examined. Alpha subband and parietal region channels were found to be most successful in predicting ADHD by using alone. An explainable model was obtained by visualizing it with gradient-weighted class activation mapping (Grad-CAM). Then, only features were extracted with the same CNN architectures and performance analyzes were carried out with different machine learning algorithms. Manual feature extraction was performed and compared with the previous stage. Decision Trees, Support Vector Machines (SVM), k-Nearest Neighborhood (k-NN) and Ensemble methods were used as classification algorithms. The most successful performance was obtained by classifying the features extracted from the AlexNet architecture with the SVM algorithm (98.33%). The high accuracy findings observed as a result of the study show that the proposed method can be used as a good decision support system in the diagnosis of ADHD.