Comparing Conventional Machine Learning and Large-Language Models for Human Stress Detection Using Social Media Posts

Mental stress has become a prevalent concern in today's fast-paced and demanding society, leading to detrimental effects on individuals' mental and physical well-being. Early detection and management of stress are crucial for maintaining a healthy lifestyle and preventing the development of chronic conditions. Stress manifests in various forms, including physiological symptoms, behavioral symptoms, Cognitive symptoms, Physical symptoms techniques have provided promising avenues for automating stress detection and monitoring processes. Earlier work has focused on wearables and physiological data, which are limited due to availability of such devices. One critical source of information relating to people's mental state, wellbeing and stress is the numerous posts made on social media. This provides a low barrier, without needing any special wearables and can provide early indicators of stress. In this paper we compare various machine learning techniques including both specifically trained models and general large language models (LLM), like OpenAI's GPT models that can be used for detecting stress based on user's social media content. We observe that the LLMs achieve very high recall (up to 99%), much higher than the best configured specifically trained conventional ML models (∼83%) with similar or higher precision and overall accuracy.