Predicting Disease Severity in Children With Attention Deficit Hyperactivity Disorder Using Dual-Branch Hypothesis Network

Attention deficit hyperactivity disorder (ADHD) is a prevalent neurodevelopmental disorder in children. Although numerous intelligent methods have been applied for its diagnosis, they seldom address symptom prediction, which is crucial for establishing the relationship between symptoms and subjective biosignals. We propose a severity prediction model, namely BSP-Net, which uses amplitude of low-frequency fluctuation (ALFF) data and constructs severity predictors within a binary hypothesis testing (BHT) framework. Specifically, we designed a dual-branch network for symptom severity prediction, with each branch corresponding to an assumed label for the test subject. Building on the accurate ADHD identification achieved by the existing auto-encoding network (AENet) within the BHT framework, we integrate its network and introduce symptom score predictors in each branch. By comparing high-level features from both branches using the AENet, we derive an estimated label. Once an assumed label is confirmed, the corresponding branch’s score predictor is activated to generate the final symptom assessment. Using the BSP-Net, our experiments achieved severity prediction accuracies of 92.4% and 84.9% on specific ADHD-200 datasets, with a score tolerance threshold of 3. Moreover, the mean squared errors on these datasets were lower than 16, significantly outperforming other methods. Importantly, discriminative brain regions corresponding to typical ALFF data in the BSP-Net were identified as ADHD biomarkers. These biomarkers align with existing research on abnormal brain regions in children with ADHD. Consequently, our method demonstrates its validity by providing biological explanations derived from ADHD mechanisms.