Novel explainable transformer model for automated attention-deficit hyperactivity disorder detection with EEG signals

Recent developments in deep learning have increased our interest in its applications in medical fields, such as the automated detection of neurodevelopmental disorders from biological signals. Attention deficit hyperactivity disorder (ADHD) is a very common condition in this group, which typically begins in childhood and often persists into adulthood. Many automated methods use an electroencephalogram (EEG) for ADHD detection. EEG signals are known to be nonlinear, non-stationary and vary significantly among individuals. To deal with this, most detection methods use either complex hand-crafted features or deep learning. The latter approach may yield better results but lacks explainability, which is critical for adoption in medical practice.

In this paper, we propose an explainable transformer-based model for automatic ADHD detection from EEG signals. To establish trust, explainability and transparency in the artificial intelligence-based approach presented, we have used Local Interpretable Model-agnostic Explanations (LIME) to visualise the model's predictions, using clinical features of EEG signals - specifically, delta (1-3 Hz), theta (4-7 Hz) and beta (14-30 Hz) components.

Our model has achieved an average accuracy of 96.2% with leave-one-out cross-validation on a public dataset. We showed that the explanations of the classification results are meaningful by linking them to clinically accepted ADHD markers.

The results demonstrate the success of the proposed explainable transformer model for ADHD detection. We hope that the proposed model has the potential to increase trust in practicing clinicians so they explore its usage as an auxiliary tool for ADHD screening in healthcare settings.