Article Link

Where we live and work, our age, and the conditions we grew up in can influence our health and lead to disparities, but these factors can be difficult for clinicians and ressearchers to capture and address. These factors, known as social determinants of health (SDoH), are often only documented in clinic notes – and therefore not readily available for automated decision-support and real-world data.

AIM researchers have demonstrated that large language models, a type of generative artificial intelligence, can be fine-tuned to automatically detect these SDoH from clinicians’ notes. These models could improve efforts to identify patient who may benefit from resource and social work support. Our fine-tuned models identified 94% of patients with an adverse SDoH, while to structured diagnostic codes identified on 2% of patients. We also showed that large language model-generated synthetic clinical text helped models more accurately sift through medical records for these “needle in a haystack” searches.

Our fine-tuned models were less prone to algorithmic bias than generalist models such as GPT-4 in zero- and few-shot settings.These new models could improve our ability to direct resources to patients who need it most, with implications for health equity. The team is planning to expand their research to additional SDoH and validate the impact of AI-identified SDoH phenotypes on patient outcomes.

This publication was featured by multiple news outlets including STAT and Politico!