Advancements in large language models (LLMs) have opened new avenues for mental health monitoring through social media analysis. In this study, we present an iterative prompt engineering framework that significantly enhances the performance of the general-purpose LLM, GPT-4, for stress detection in social media posts, leveraging psychologist-informed hints. This approach achieved a substantial 17% accuracy improvement from 72% to 89% for the January 2025 version of GPT-4, alongside an 80% reduction in false positives compared to baseline zero-shot prompting. Our method not only surpassed domain-specific models like Mental-RoBERTa by 5% but also uniquely generates human-readable rationales. These rationales are crucial for mental health professionals, assisting them in understanding and validating the model’s outputs—a key benefit for sensitive mental health applications. These results highlight prompt engineering as a resource-efficient, transparent strategy to adapt general-purpose LLMs for specialized tasks, offering a scalable solution for mental health monitoring without the need for costly fine-tuning.

Accurate and secure classifying informal documents related to mental disorders is challenging due to factors such as informal language, noisy data, cultural differences, personal information and mixed emotions. Conventional deep learning models often struggle to capture patterns in informal text, as they miss long-range dependencies, explain words and phrases literally, and have difficulty processing non-standard inputs like emojis. To address these limitations, we expand data dimensionality, transforming and fusing textual data and signs from a 1D to a 2D space. This enables the use of pre-trained 2D CNN models, such as AlexNet, Restnet-50, and VGG-16 removing the need to design and train new models from scratch. We apply this approach to a dataset of social media posts to classify informal documents as either related to suicide or non-suicide content. Our results demonstrate high classification accuracy, exceeding 99%. In addition, our 2D visual data representation conceals individual private information and helps explainability.