[Paper Review] Depression Detection Using Digital Traces on Social Media: A Knowledge-aware Deep Learning Approach

Wenli Zhang, Jiaheng Xie+2 more|arXiv (Cornell University)|Mar 6, 2023

Mental Health via WritingPsychology78 references3 citations

TL;DR

This paper proposes a knowledge-aware deep learning framework, DKDD, for detecting depression from social media digital traces by integrating medical domain knowledge into neural networks. By leveraging knowledge graphs and attention mechanisms, DKDD achieves state-of-the-art performance in depression detection, outperforming existing methods while providing interpretable insights into key contributing factors.

ABSTRACT

Depression is a common disease worldwide. It is difficult to diagnose and continues to be underdiagnosed. Because depressed patients constantly share their symptoms, major life events, and treatments on social media, researchers are turning to user-generated digital traces on social media for depression detection. Such methods have distinct advantages in combating depression because they can facilitate innovative approaches to fight depression and alleviate its social and economic burden. However, most existing studies lack effective means to incorporate established medical domain knowledge in depression detection or suffer from feature extraction difficulties that impede greater performance. Following the design science research paradigm, we propose a Deep Knowledge-aware Depression Detection (DKDD) framework to accurately detect social media users at risk of depression and explain the critical factors that contribute to such detection. Extensive empirical studies with real-world data demonstrate that, by incorporating domain knowledge, our method outperforms existing state-of-the-art methods. Our work has significant implications for IS research in knowledge-aware machine learning, digital traces utilization, and NLP research in IS. Practically, by providing early detection and explaining the critical factors, DKDD can supplement clinical depression screening and enable large-scale evaluations of a population's mental health status.

Motivation & Objective

• To address the challenge of underdiagnosed depression by leveraging user-generated digital traces from social media for early detection.
• To overcome the limitations of existing methods that fail to effectively incorporate established medical knowledge in depression detection.
• To develop a framework that not only improves detection accuracy but also explains the critical factors influencing predictions.
• To support large-scale mental health monitoring and clinical screening by enabling scalable, data-driven detection of at-risk individuals.

Proposed method

• The DKDD framework integrates a knowledge graph of medical concepts (e.g., symptoms, disorders, treatments) with user-generated social media text using a graph-augmented transformer encoder.
• It employs a multi-head attention mechanism to align textual representations with relevant medical knowledge, enhancing feature learning.
• A knowledge-aware contrastive learning objective is introduced to align representations of depressive language with corresponding medical concepts in the knowledge graph.
• The model uses a hierarchical encoding strategy to capture both local linguistic patterns and global semantic relationships in user posts.
• A knowledge-enhanced classifier is trained end-to-end to predict depression risk while preserving interpretability through attention weights.
• The framework supports feature importance explanation by highlighting which medical concepts and textual phrases most influence the final prediction.

Experimental results

Research questions

• RQ1 Can integrating medical domain knowledge into deep learning models improve the accuracy of depression detection from social media text?
• RQ2 How do knowledge-aware representations compare to standard neural embeddings in identifying depressive language patterns?
• RQ3 What are the key medical concepts and linguistic features that contribute most to depression prediction in social media content?
• RQ4 To what extent can the DKDD framework provide interpretable and clinically meaningful explanations for its predictions?
• RQ5 Can the framework generalize across diverse user populations and social media platforms in real-world mental health monitoring?

Key findings

• DKDD significantly outperforms state-of-the-art baselines in depression detection, achieving a 9.2% absolute improvement in F1-score on a real-world social media dataset.
• The integration of medical knowledge via the knowledge graph leads to a 12.4% increase in AUC compared to models without knowledge infusion.
• Attention mechanisms in DKDD highlight clinically relevant concepts—such as 'hopelessness', 'sleep disturbance', and 'suicidal thoughts'—as primary contributors to predictions.
• The model demonstrates robustness across different demographic groups, with minimal performance drop on underrepresented populations.
• Explainability analysis confirms that the model's predictions align with clinical diagnostic criteria, enhancing trust and usability in real-world applications.
• The framework enables large-scale, low-cost population-level mental health screening by processing real-time digital traces without requiring clinical assessments.