Evaluating Large Language Models for Extracting Social Determinants of Health in Substance Use Disorder Notes

Automated extraction of social determinants of health (SDoH) may support earlier identification of unmet social needs and inform substance use disorder (SUD) care. This study provides preliminary evidence that few-shot prompting improves large language model performance in extracting SDoH from discharge summaries of patients with SUD. While the accuracy of few-shot prompting exceeded zero-shot prompting, limitations such as low inter-annotator agreement and hallucinations remain. Future work should explore domain-adapted models, active learning, and validation in diverse populations.