Improving Suicidal Ideation Detection in Social Media Posts: Topic Modeling and Synthetic Data Augmentation Approach.

backgroundIn an era dominated by social media conversations, it is pivotal to comprehend how suicide, a critical public health issue, is discussed online. Discussions around suicide often highlight a range of topics, such as mental health challenges, relationship conflicts, and financial distress. However, certain sensitive issues, like those affecting marginalized communities, may be underrepresented in these discussions. This underrepresentation is a critical issue to investigate because it is mainly associated with underserved demographics (eg, racial and sexual minorities), and models trained on such data will underperform on such topics.

objectiveThe objective of this study was to bridge the gap between established psychology literature on suicidal ideation and social media data by analyzing the topics discussed online. Additionally, by generating synthetic data, we aimed to ensure that datasets used for training classifiers have high coverage of critical risk factors to address and adequately represent underrepresented or misrepresented topics. This approach enhances both the quality and diversity of the data used for detecting suicidal ideation in social media conversations.

methodsWe first performed unsupervised topic modeling to analyze suicide-related data from social media and identify the most frequently discussed topics within the dataset. Next, we conducted a scoping review of established psychology literature to identify core risk factors associated with suicide. Using these identified risk factors, we then performed guided topic modeling on the social media dataset to evaluate the presence and coverage of these factors. After identifying topic biases and gaps in the dataset, we explored the use of generative large language models to create topic-diverse synthetic data for augmentation. Finally, the synthetic dataset was evaluated for readability, complexity, topic diversity, and utility in training machine learning classifiers compared to real-world datasets.

resultsOur study found that several critical suicide-related topics, particularly those concerning marginalized communities and racism, were significantly underrepresented in the real-world social media data. The introduction of synthetic data, generated using GPT-3.5 Turbo, and the augmented dataset improved topic diversity. The synthetic dataset showed levels of readability and complexity comparable to those of real data. Furthermore, the incorporation of the augmented dataset in fine-tuning classifiers enhanced their ability to detect suicidal ideation, with the F

conclusionsOur results demonstrate that synthetic datasets can be useful to obtain an enriched understanding of online suicide discussions as well as build more accurate machine learning models for suicidal narrative detection on social media.