Architectures and strategies for the differential diagnosis of adolescent depression and anxiety: bridging tabular deep learning, interpretability, and scale streamlining

Escalating societal and occupational pressures have exacerbated the global mental health crisis. The symptomatic overlap between depression and anxiety presents a formidable diagnostic hurdle for differential diagnosis in mental health. developing AI-driven diagnostic models to differentiate these states and streamlining complex clinical instruments is of substantial clinical and scientific value.

Utilizing clinical questionnaires and physician-confirmed diagnoses, distinct datasets were established. This study evaluates four dimensions: model architecture, data strategy optimization, interpretability, and clinical instrument streamlining. Four resampling strategies and tabular generative models were integrated to address data imbalance. Through stratified cross-validation, 29 models-spanning traditional Machine Learning (ML), sequential Deep Learning (DL), and advanced tabular DL-were compared. Optimization strategies focusing on input representation and generalization were also evaluated.

Data strategies significantly enhanced standard deep learning architectures. Specifically, Mambular-TTT combined with Borderline-SMOTE and TVAE achieved a 7.3% relative AUC improvement over the best-performing traditional machine learning baseline. Hierarchical SHAP identified childhood trauma and neuroticism as shared correlated predictive features while revealing subtype-specific feature association patterns. Stepwise reduction revealed that a 7-instrument subset optimally balances diagnostic performance and clinical efficiency; the Area Under the Receiver Operating Characteristic Curve (AUC) for the streamlined Dataset 2 improved to 0.70, outperforming the full-scale model.

This study demonstrates that reliable AI-assisted diagnosis requires the integrated optimization of model architecture and data strategy. This approach, combined with an interpretability-based scale streamlining strategy, provides a practical and evidence-based implementation path for transforming machine learning models into efficient, deployable clinical diagnostic tools.