SynthEHR-eviction: enhancing eviction SDoH detection with LLM-augmented synthetic EHR data

Comprehensive Summary

This study introduces SynthEHR-Eviction, a scalable pipeline that uses large language models to detect eviction-related social determinants of health from unstructured electronic health records. Because eviction is rarely coded in structured medical data, the authors combine human-in-the-loop annotation, automated prompt optimization, and reasoning-based fine-tuning to generate high-quality synthetic training data. The resulting dataset includes 14 detailed SDoH categories and significantly improves model performance. Fine-tuned LLMs (such as Qwen2.5 and LLaMA3) achieved high accuracy (Macro-F1 up to around 90%), outperforming baseline models like GPT-4o and BioBERT, while reducing annotation effort by over 80%.

Outcomes and Implications

This work improves the ability to identify housing instability and eviction from clinical notes, enabling better integration of social determinants into healthcare decision-making. It supports more proactive interventions for at-risk patients and demonstrates how synthetic data and LLMs can scale public health research. The approach could also be generalized to detect other underreported social factors in healthcare data, ultimately helping address health disparities.