DRNet: Early Recognition of Depression Based on National Health Survey Data

Depression is a leading cause of health loss and disability worldwide. Early recognition is essential for the treatment of depression. Research on automatic depression recognition methods is crucial. The depression recognition methods based on structured personal features can achieve high accuracy. But these methods lack an explanation of how these features play a role in depression recognition. Our research develop a deep learning model termed DRNet for early depression recognition (DR). The input data is from national health survey dataset. The model focuses both on low-order linear relations and high-order cross-relations among features, which are important for DR. Based on DRNet, the roles features play in DR can be explained in the form of contribution. Experimental results show that DRNet can accurately discriminate depression patients with the best AUC and F1 Score in comparison with some machine learning models (Logistic Regression, Random Forest, and Extreme Gradient Boosting). In the results of feature contribution analysis, almost all features with high contribution to DR can be directly or indirectly verified in previous studies on related factors of depression. In conclusion, an efficient and interpretable deep learning model for DR is proposed, which contributes to the auxiliary diagnosis and the discovery of potential risk factors for depression.