Combination of ConvLSTM and Attention mechanism to diagnose ADHD based on EEG signals

Abstract

Among neurodevelopmental disorders, attention deficit hyperactivity disorder (ADHD) is the most prevalent disorder in childhood. Early diagnosis and treatment of this disorder can reduce negative impacts, such as learning difficulty, antisocial behaviours, financial problems and divorce in adulthood. Although clinical diagnoses are currently available, they are based on patient behaviours and are not reliable. Researchers developed different methods to discover a biomarker that can help accurate diagnosis. Biological signals such as electroencephalography (EEG) draw the most interest because of their ability to record neurons electrical activity. We propose a deep learning framework that combines the ConvLSTM and attention mechanism. To provide the input for this framework, we first calculate a dynamic connectivity tensor. This technique is more effective than feature extraction methods such as Fourier transform-based approaches and nonlinear analyses. Due to the structure of ConvLSTM, the model can extract temporal and spatial features simultaneously, and the attention mechanism provides insights for the model to score different time instants in EEG data. These two steps lead to effectively encoding a compact representation of EEG signals. It is the first time to apply ConvLSTM and the attention mechanism combination on time series data. To examine the proposed framework, we run our experiments on 400 data instances. We trained our model using 5-fold cross-validation. After ten different executions, the best model has an accuracy of 99.75%, which is the superior performance among the studies on this data.