Multi-LSTM Networks for Accurate Classification of Attention Deficit Hyperactivity Disorder from Resting-State fMRI Data

Attention deficit hyperactivity disorder (ADHD) is a widespread mental disorder among young children. Due to the complex pathological mechanisms and clinical symptoms, the diagnosis of ADHD is still challenging. In this paper, we propose a novel multi-network of long short term memory (multi-LSTM) for the identification of ADHD. The Gaussian mixture model (GMM) is introduced to cluster different regions of interests (ROIs) for feature selection. Then, the data augmentation and phenotypic information are used to further improve the classification performance. The simulation experiment demonstrates that the proposed model outperforms the state-of-the-art methods based on the multi-site ADHD-200 global competition dataset. It is anticipated that the proposed ROI-based clustering method and multi-LSTM model can provide valuable insights into the auxiliary diagnosis of ADHD from the rs-fMRI signal.