Study compares four machine learning methods for detecting caregiver burden

[medRxiv] A preprint study compared four machine learning and AI approaches for detecting dementia caregiver burden in de-identified caregiver-clinician transcripts, using Zarit burden scores as reference. According to the authors, Klein et al., in transcripts from the COPE in PACE study a naive baseline predicting burden in every transcript achieved precision/F1 of 0.40/0.57, LLM prompting 0.43/0.60 and weak supervision 0.48/0.58; a fully supervised model trained on sentence-level labels with only 1.1% burden-positive sentences reached precision/F1 of 0.19/0.25, a result the authors said is not directly comparable because of different labels and units of analysis. The authors said LLM prompting modestly outperformed weak supervision and the naive baseline without fine-tuning, while class imbalance limited fully supervised sentence-level classification, and that larger annotated corpora and refined labeling strategies are needed for clinical deployment.