Hierarchical Deep Neural Network for Child Brain Health Assessment

Early assessment of mental health in youth is vital for prevention and timely intervention, yet current clinical practice relies on subjective, labor-intensive questionnaires. This study presents a hierarchical deep neural network that predicts dimensional psychopathology scores from electroencephalography (EEG). Task-specific encoders are trained on power spectral density (PSD) and Hjorth features, demographics, and task annotations, and a second-stage Long Short-Term Memory (LSTM) fuses per task embeddings to produce subject-level scores. Using the Healthy Brain Network EEG (HBN-EEG) dataset, evaluation occurs under a leave-one-release-out protocol with 5-fold cross-validation. The proposed model achieves a higher R2 and lower root mean square error (RMSE) than the baselines with statistically significant gains in most comparisons, and shows robustness to heterogeneity in task completion, run counts, and session durations. These results demonstrate considerable robustness against dataset subject- and task-level heterogeneity. This study highlights great potential for an effective and comprehensive AI-driven evaluation method for mental healthcare.