Comparing Machine Learning Methods for Detecting Caregiver Burden in Transcribed Conversations with Clinicians

LLM prompting on de-identified caregiver–clinician transcripts from the COPE in PACE study detects caregiver burden with precision/F1 of 0.43/0.60, modestly outperforming non-fine-tuned baselines.

• Why it matters: Elevated caregiver burden is linked to earlier patient institutionalization and higher caregiver morbidity, but routine encounters lack a standardized, objective way to detect it.
• What they did: Four AI approaches were compared using Zarit burden scores on COPE in PACE de-identified transcripts, including a naive baseline, LLM prompting, weak supervision, and a fully supervised sentence-level model.
• The result: LLM prompting modestly outperformed the naive baseline and weak supervision without fine-tuning, while fully supervised sentence-level classification was constrained by class imbalance (1.1% burden-positive sentences).

AI summary of the abstract · may contain errors · not medical advice · read the paper.