A Hybrid Deep Learning Framework with Swarm-Optimized Ensemble Learning for Depression Detection from Social Media Text

Abstract

Depression is a pervasive mental health disorder and a leading contributor to global disability, often remaining undiagnosed due to social stigma and the limitations of traditional screening methods. The widespread use of social media platforms generates a rich corpus of textual data that can serve as a passive indicator of an individual’s mental state. This paper proposes a novel hybrid deep learning framework for the detection of depressive symptoms from social media text. The framework integrates a Convolutional Neural Network (CNN) for extracting local linguistic features with a Bidirectional Long Short‑Term Memory (BiLSTM) network for capturing long‑range contextual dependencies. To further enhance performance, the model is integrated into an ensemble learning architecture, where the hyperparameters of the individual models are optimized using a Particle Swarm Optimization (PSO) algorithm. The proposed methodology is evaluated on a publicly available dataset of Reddit posts annotated for depressive content. Experimental results demonstrate that the proposed hybrid ensemble model, optimized via PSO, achieves superior performance in terms of accuracy, precision, recall, and F1‑score compared to several baseline machine learning and standard deep learning models. The findings underscore the efficacy of combining feature extraction techniques with optimized ensemble strategies for developing robust, automated depression detection systems.