Abstract
Background: Depression and anxiety are common and often detected late. Deep learning needs large labelled datasets, which are rare in mental health. Transfer learning reuses knowledge from pre-trained models and may reduce this problem.
Objective: To map how deep transfer learning has been used to detect, classify or estimate the severity of depression and anxiety, and to examine how the reported results depend on validation design.
Methods: The review followed PRISMA 2020. PubMed and OpenAlex were searched on 27 October 2026 for studies published from January 2019 to September 2026. The earlier version of this manuscript was also checked for eligible studies. Primary studies were included if they applied a pre-trained deep model, fine-tuning, domain adaptation, self-supervised pre-training or a large language model to individual-level detection or severity estimation. Data were extracted from full abstracts and synthesized narratively.
Results: Of 798 records, 318 studies met the criteria. Most studies (81.1%) were published from 2024 onward. Depression alone was studied in 288 studies, anxiety alone in 10 and both in 20. Text from social media (34.9%) and clinical interviews or narratives (17.3%) were the main data sources. Fine-tuning (143 studies) and feature extraction (86) were most common, and 87 studies used large language or foundation models. Only 45 studies (14.2%) reported participant-independent or external validation, and 98 abstracts gave no numerical result. The median reported accuracy was 91.3%, and external validation gave lower values. Small local language models performed poorly on interview transcripts.
Conclusions: Transfer learning is now the default approach in this field, but the evidence is mostly based on random splits of small or public benchmark datasets. Reported accuracies should not be read as clinical performance. Future work needs participant-independent and external testing, TRIPOD+AI reporting, more work on anxiety and comorbidity, and data from under-represented languages such as Arabic.
Keywords
Depression, Anxiety, Transfer learning, Deep learning, Large language models, Pre-trained models, Speech, Social media, Mental health screening, Systematic review