ADHD classification using auto-encoding neural network and binary hypothesis testing

Area of Science:

• Neuroscience
• Computer Science
• Medical Imaging

Background:

• Attention Deficit Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder in school-aged children.
• Accurate and early diagnosis of ADHD is critical for effective treatment.
• Existing ADHD classification methods face challenges with insufficient data and feature noise from related disorders.

Purpose of the Study:

• To propose a novel deep-learning classification architecture for ADHD.
• To overcome limitations of insufficient data and feature noise in ADHD diagnosis.
• To enhance the objectivity and reliability of neurobiological ADHD classification.

Main Methods:

• A deep-learning architecture combining a binary hypothesis testing framework and a modified auto-encoding (AE) network.
• Utilizing brain functional connectivities (FCs) from both training and test data for feature selection.
• Employing the AE network to capture effective features and reduce inter- and intra-class variability disturbances.

Main Results:

• The proposed method significantly outperforms existing ADHD classification techniques.
• Achieved an average accuracy of 99.6% using leave-one-out cross-validation on the ADHD-200 database.
• Demonstrated robustness and practical convenience with uniform parameter settings across datasets.

Conclusions:

• The novel deep-learning approach effectively addresses data scarcity and feature noise in ADHD classification.
• The method provides a highly accurate, robust, and convenient tool for neurobiological ADHD diagnosis.
• This advancement holds significant potential for improving clinical ADHD assessment and management.