Purpose – This systematic literature review examines recent approaches to depression detection from social media, with particular emphasis on transformer-based text representation, deep autoencoders for dimensionality reduction, and metaheuristic optimisation for feature selection and hyperparameter tuning.Methods – Following PRISMA 2020 reporting guidelines, the review synthesised 85 peer-reviewed studies published between 2018 and 2026 across six databases. The analysis focused on transformer-based embeddings, deep autoencoders, and hybrid metaheuristic optimisation techniques used for depression detection from social media data.Findings – BERTweet demonstrates strong capability in capturing Twitter-specific linguistic nuances, while deep convolutional autoencoders reduce high-dimensional embeddings into compact latent representations. Hybrid metaheuristic algorithms, particularly adaptive PSO-GWO frameworks, offer advantages in addressing the exploration-exploitation trade-off. However, existing studies rarely incorporate adaptive balance mechanisms that dynamically adjust optimisation behaviour using population diversity and solution-improvement indicators.Originality/Value – Unlike previous reviews that mainly catalogue predictive performance, this review identifies adaptive optimisation balance as a critical unresolved methodological gap. It also proposes a conceptual integration of transformer representations, autoencoder-based dimensionality reduction, and adaptive hybrid metaheuristic optimisation.Research Implications – Future studies should empirically validate integrated frameworks combining these components and extend them through multimodal fusion, explainable AI, and federated learning. The review also identifies exploratory implications for AI literacy education, although these remain conceptual rather than empirically established.
Copyrights © 2026