DepressLLM: Interpretable Domain-Adapted Language Model for Depression Detection from Real-World Narratives

Advances in large language models (LLMs) have enabled a wide range of applications. However, depression prediction is hindered by the lack of large-scale, high-quality, and rigorously annotated datasets. In this study, a novel corpus of 3,699 autobiographical narratives reflecting both happiness and distress was used to develop DepressLLM. DepressLLM generates interpretable predictions via the proposed Score-guided Token Probability Summation (SToPS) module. This module not only improves the model's classification performance, achieving an AUC of 0.789 that rises to 0.904 on samples with 95% confidence, but it also delivers reliable confidence estimates. We evaluated DepressLLM on an in-house ecological momentary assessment corpus and multiple public interview datasets, demonstrating robust performance across diverse testing settings. We further analyzed errors across the test set and conducted a psychiatrist review of high-confidence misclassifications to identify key limitations and inform future refinements. These findings demonstrate that interpretable artificial intelligence (AI) can enable early diagnosis of depression and underscore the potential of medical AI in psychiatry.