Guided Clinical Reasoning with LLMs: Reconstructing the PHQ-9 from Multimodal Inputs
Conference
Trujillo, R, Poellabauer, C. (2026). Guided Clinical Reasoning with LLMs: Reconstructing the PHQ-9 from Multimodal Inputs
. 149-158. 10.1109/ICHI69079.2026.00030
Trujillo, R, Poellabauer, C. (2026). Guided Clinical Reasoning with LLMs: Reconstructing the PHQ-9 from Multimodal Inputs
. 149-158. 10.1109/ICHI69079.2026.00030
This work presents a zero-shot, symptom-level inference framework in which general-purpose large language models (LLMs) infer individual PHQ-9 depressive symptoms from multimodal behavioral data, including physiological signals, emotional self-reports, and narrative transcripts. Using the MERSA dataset, which includes Fitbit metrics (e.g., heart rate, sleep, circadian rhythm), daily PANAS surveys, and short self-narratives, we evaluate this framework by prompting several prompting several LLMs from the GPT and DeepSeek model families to infer PHQ-9 symptoms. Ablation trials and prompt variations are used to assess the contribution of each modality and the robustness of inference across heterogeneous inputs. GPT-4o mini and DeepSeek-R1 achieved strong performance (mean F1 scores of 0.76), particularly for somatic symptoms such as fatigue and sleep disturbance (F1≈0.9), while suicidal ideation and appetite change were not reliably inferred. The models differed in precision-recall trade-offs, with GPT-4o mini favoring precision and DeepSeek-R1 favoring recall. In ablations, models (especially GPT-4o mini) retained the ability to infer some psychological symptoms from wearable-derived features and some somatic symptoms from self-reports alone, consistent with cross-domain symptom inference. Input pipeline variations had minimal impact and low-information modalities (e.g., transcripts) did not substantially degrade performance. In general, the findings suggest that general-purpose LLMs can support flexible, symptom-level mental health inference over structured and partially missing data without task-specific fine-tuning, with model selection guided by application-specific recall-precision requirements.
