Classification of familial and non-familial ADHD using auto-encoding network and binary hypothesis testing.
Authors
Affiliations (7)
Affiliations (7)
- Department of Biomedical Engineering, New Jersey Institute of Technology, NJ, USA; Graduate School of Biomedical Sciences, Rutgers University, Newark, NJ, USA.
- Department of Biomedical Engineering, New Jersey Institute of Technology, NJ, USA.
- Department of Rehabilitation and Movement Sciences, Rutgers University, Newark, NJ, USA.
- School of Computing, University of Georgia, Athens, GA, USA.
- School of Biomedical Sciences and Engineering, South China University of Technology, Guangzhou International Campus, Guangzhou, China.
- Department of Psychology, Brooklyn College, the City University of New York, Brooklyn, New York, USA.
- Department of Biomedical Engineering, New Jersey Institute of Technology, NJ, USA; Department of Electrical and Computer Engineering, New Jersey Institute of Technology, NJ, USA. Electronic address: [email protected].
Abstract
Family heritage is one of the most powerful risk factors for attention-deficit/hyperactivity disorder (ADHD). Children with familial ADHD (ADHD-F) and non-familial ADHD (ADHD-NF) have both shared and distinct behavioral characteristics and clinical outcomes, with their neuropathological underpinnings under-investigated. In this study, we utilized an autoencoder-based deep learning architecture within the binary hypothesis framework to identify the most important structural- and diffusion-MRI-based neural signatures that robustly discriminate 129 children with ADHD-F, 159 AHDH-NF, and 150 matched controls. Nested leave-one-out and five-fold methods were used for double cross-validations of the results. Classification accuracy, sensitivity, specificity, and area under the curve (AUC) were applied to evaluate the model performance. The model achieved accuracies of 65.4 ± 3.2, 67.0 ± 2.2, and 62.0 ± 4.9, with corresponding AUCs of 65.6 ± 4.8, 67.6 ± 5.2, and 65.8 ± 5.2, for ADHD-F vs. controls, ADHD-NF vs. controls, and ADHD-F vs. ADHD-NF, respectively. The most informative features for successful ADHD-F vs. control discrimination were mean diffusivity (MD) of right fornix, MD of left parahippocampal cingulum, and cortical thickness of right inferior parietal cortex. The key contributors for successful ADHD-NF vs. control discrimination were fractional anisotropy (FA) of left inferior fronto-occipital fasciculus, MD of right fornix, and cortical thickness of right medial orbitofrontal cortex. The highlighted features for successful ADHD-F vs. ADHD-NF discrimination were volume of left cingulate cingulum tract, volume of right parietal segment of the superior longitudinal fasciculus, and cortical thickness of right fusiform cortex. Our binary hypothesis semi-supervised deep learning framework robustly discriminated familial vs. non-familial ADHD and provided validated neural features that have potential to serve as distinct treatment targets of ADHD-F vs. ADHD-NF.