RST-Enhanced Depression Detection: A Feature-Fusion Ensemble Framework

Abstract

Early identification of depression risk from social media text can support large-scale screening and timely follow-up. However, posts are often emotionally complex and linguistically ambiguous, which makes robust detection challenging. This paper proposes RST-DS (Rhetorical Structure Theory-based Depression Scanning), a discourse-aware framework that integrates Rhetorical Structure Theory (RST) signals with lexical evidence for classifying posts as Depressed/Non-Depressed (D/ND). Using Reddit posts, we compute an RST-derived score capturing rhetorical relationships and coherence and fuse it with lexical features represented via Term Frequency–Inverse Document Frequency using two strategies: feature addition (+) and feature concatenation (||). We evaluate K-Nearest Neighbors (KNN), Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP), along with a soft-voting ensemble model named MLPBoostReg (an ensemble of LR, XGBoost, and MLP). Using a 5-fold cross validation, the concatenation strategy consistently outperforms the addition strategy across accuracy, precision, recall, and F1-score. The best-performing configuration, MLPBoostReg|| achieves the highest accuracy, precision, recall, and F1-score of 0.970, indicating a strong balance between identifying depression-related content and limiting false alarms. These findings suggest that combining rhetorical structure with lexical evidence improves depression detection from social media text, while noting that social-media labels do not substitute for clinical diagnosis.