Hybrid CNN–BiLSTM for EEG-Based ADHD Detection in Children

Attention Deficit Hyperactivity Disorder (ADHD) is a common neuropsychological condition that affects attention, behavior, and learning in kids. The conventional method for diagnosis includes clinical interview and behavior ratings scales; however, this approach lacks objectivity. In this regard, this paper presents a novel hybrid architecture of deep learning combining one-dimensional CNN with Bidirectional LSTM networks for detection of ADHD through EEG recordings. A freely available database from IEEE Data Port was used, containing resting state recording data of 121 children in the age group 7-12 years, where 61 were diagnosed with ADHD and 60 were non-ADHD patients, recorded at 128Hz on 19 electrode placements on the scalp. The pre-processing step included band-pass filter, notch filter, removal of artifacts through independent component analysis, and segmentation of epochs of constant size. The CNN layer learns the short-range temporal and spatial features, while the BiLSTM layer learns the long-range temporal dependencies. Cross validation by ten folds per subject was applied in order to validate the model. It was the most precise (97.3%), sensitive (97.7%), specific (96.8%), with the highest F1-score (0.971) and area under the curve (AUC) (0.989) compared to classical machine learning algorithms, CNNs, LSTMs and hybrid CNNLSTM models. The obtained results indicated that modeling simultaneously both space and time bidirectionally is more efficient in predicting ADHD from EEG data compared to single stream approaches, and can be considered a non-invasive test for ADHD.