Text-Based Depression Models Demonstrate Strong Predictive Performance

Authors systematically reviewed text-based depression estimation studies (searching 2014–2025) and meta-analyzed 15 models from 11 studies, reporting a pooled correlation r=0.605. Embedding-based features and deep-learning architectures (r≈0.74 and r≈0.73) outperformed traditional features and shallow models, and clinician-diagnosed labels yielded higher performance than self-reports; reporting quality was positively associated with performance.