Search papers, labs, and topics across Lattice.
This study introduces a two-stage framework for sentence-level depression symptom recognition that separates symptom candidate generation from definition-guided verification. Utilizing a contrastively fine-tuned sentence encoder for candidate generation and a fine-tuned language model for verification against diagnostic definitions, the approach outperforms existing methods, achieving the highest accuracy and F1 scores. The findings suggest that decomposing the recognition process enhances performance, although challenges remain for less common symptom categories.
Achieving superior accuracy in depression symptom recognition hinges on a novel two-stage framework that combines candidate generation with definition-guided verification.
Sentence-level recognition of depression symptoms is challenging because similar expressions can differ in symptom relevance, and language-model inference is insufficiently grounded in diagnostic definitions. This study proposes a two-stage framework separating symptom-candidate generation from definition-grounded verification. A contrastively fine-tuned sentence encoder generates a symptom candidate per sentence, and a fine-tuned language model verifies whether the candidate is present or absent using the sentence, its context, and a candidate-specific diagnostic definition, checking its judgment against that definition before answering. Evaluated against encoder, inference-based, medical, and general LLM baselines and a matched single-stage supervised classifier, the proposed pipeline attains the best accuracy and F1 scores of all methods, with rationales matching expert-authored annotations. A preliminary clinical audit indicates moderate alignment with diagnostic definitions, with explanation quality strongly dependent on prediction correctness. The results support decomposing symptom recognition into candidate generation and definition-grounded verification, though performance remains limited for rare categories.
