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

Dementia caregivers' elevated burden is associated with earlier patient institutionalization and greater caregiver morbidity, yet no standardized, objective method exists for detecting burden during routine clinical encounters. We compared four machine learning and AI approaches for detecting caregiver burden in de-identified caregiver-clinician transcripts from the COPE in PACE study using Zarit burden scores. A naive baseline predicting burden in every transcript achieved precision/F1 = 0.40/0.57. LLM prompting achieved precision/F1 = 0.43/0.60. Weak supervision achieved precision/F1 = 0.48/0.58. The fully supervised model, trained on sentence-level labels with only 1.1% burden-positive sentences, achieved precision/F1=0.19/0.25; this result is not directly comparable because of different labels and units of analysis. LLM prompting modestly outperformed weak supervision and naive baseline without fine-tuning, while class imbalance limited fully supervised sentence-level classification. Larger annotated corpora and refined labeling strategies are needed for clinical deployment.