Predicting PHQ-9 Depression Scores from Free-Text Responses Using a Hybrid SBERT and Bi-LSTM architecture

Abstract

The Patient Health Questionnaire-9 (PHQ-9) is a clinically validated tool for assessing depression severity across symptoms such as mood, energy, sleep, and appetite. However, standard PHQ-9 administration relies on fixed categorical response options, which may not capture the nuance and contextual detail present in natural language descriptions of symptoms. This study proposes a hybrid, transfer-learning approach to predict continuous PHQ-9 total scores from 250 students’ semi-structured free-text responses. Our model uses a domain-adapted transformer backbone within the Sentence-Transformers framework to encode each PHQ-9 item response into a dense embedding, and feeds the resulting nine-embedding sequence into a bidirectional LSTM regressor to capture cross-item dependencies. The encoder is fine-tuned on response pairs to better align the embedding space with PHQ-9–related semantics. Evaluated using k-fold cross-validation, the proposed model achieves an average MAE of 1.59, and attains over 97% accuracy after mapping predicted scores to standard PHQ-9 severity categories, outperforming multiple transformer-based encoder baselines and showing more consistent performance across folds. These findings suggest that NLP-based modeling can support PHQ-9 screening by estimating severity scores from natural-language responses in student populations, complementing conventional questionnaire scoring.