Comparison of Depression Detection Between LLMs and Zero-Shot Learning Using DAD Dataset

Early and accurate detection of depression is critical for improving mental health outcomes. Yet, selecting an appropriate machine learning approach remains challenging due to differences in data availability, annotation costs, and computational resources. Fine-tuned Large Language Models (LLMs) and Zero-Shot Learning (ZSL) represent fundamentally distinct paradigms for natural language understanding tasks. Comparing these approaches is crucial because LLMs rely on large, labeled datasets, while ZSL provides a practical alternative in scenarios with limited or no annotated data. Despite their growing application, these methods are often studied in isolation, leaving a gap in understanding their relative advantages and trade-offs. This study addresses this gap by evaluating the performance of a fine-tuned BERT-LLM and a BART-based ZSL model on our custom-made Depression Analysis Dataset (DAD). Our research experiments showed that the fine-tuned BERT-LLM achieved accuracy and F1 scores of 98.94% and 99.24 %, respectively, while the BART-ZSL model attained 76.53% and 81.26%. The findings highlight the suitability of LLMs for high-reliability settings and ZSL for early detection in resource-constrained contexts, providing actionable insights for data scientists and mental health practitioners to design effective and accessible depression detection systems.