Automatically detecting trends and open questions from mental health publications: a Wellcome-funded GALENOS project.

BACKGROUND: More effective and better tolerated treatments are urgently needed for people with mental health disorders, such as anxiety, depression and psychosis. However, the rate of translation of positive results from early phase studies into clinically validated treatments remains painstakingly slow. The scientific literature on mental health preclinical and early interventions is burgeoning at pace, making it difficult for researchers, practitioners and policymakers to identify and track new developments. OBJECTIVE: As part of the Wellcome-funded Global Alliance of Living Evidence for aNxiety, depressiOn and pSychosis project, we aimed to develop and evaluate an automated approach to track the evolution of mental health research over time, detect emerging trends and suggest open questions. METHODS: Our approach used topic modelling, large language models and time-series forecasting in combination. We applied our approach to a corpus of 182 747 titles and abstracts extracted from the OpenAlex database for 2015-2025. Using topic modelling to identify topics and then tracking topic mentions over time, we built a time series predictive model and predicted 'trendiness' based on sustained increased mentions above baseline expected from model predictions. We evaluated our approach retrospectively using a blinded expert study of a randomly selected sample of trending and not trending topics. Finally, we developed a novel topic-augmented generation approach to suggest open questions in trendy topics and evaluated the approach by comparison to baseline-generated questions without topic augmentation. FINDINGS: Our approach detected 973 topics and predicted 165 (17%) of those as trending. Key topics that the model predicted as trending included 'ketamine for treatment-resistant depression', 'student mental health in academia' and 'COVID-19 psychosis'. We found that domain experts largely agreed with the model's predictions of trendiness. Topic-augmented generated questions were more specific than baseline generated questions. CONCLUSIONS: Our approach enables identification of new developments and open questions. Future work will improve temporal pattern tracking and use full texts. CLINICAL IMPLICATIONS: Our approach can support all stakeholders to gain an overview of the published literature, assess temporal patterns, identify trends and rank open questions.