AI model trained to identify patients’ social circumstances

Editor's Note: 

Large language models trained to extract patients’ social determinants of health (SDoH) from clinician notes could help to identify patients who need additional support and resources. The findings, from investigators at Mass General Brigham, appeared in the Nature journal Digital Medicine on January 11. 

Housing circumstances, employment, access to transportation, and other SDoH can have a major impact on health outcomes, but doctors using traditional diagnostic tools have access to this information in only 2% of cases, the researchers note.

For the study, they manually reviewed 800 clinician notes from 770 patients with cancer and tagged sentences that referred to six pre-determined SDoH: employment status, housing, transportation, parental status, relationships, and presence or absence of social support. They then trained a language model to identify references to SDoH in clinician notes, testing the model on an additional 400 clinic notes from patients treated with immunotherapy. The resulting language model could consistently identify rare references to SDoH in clinician notes. 

The researchers say their goal is better identify patients who can benefit from social work support and additional resources and call attention to the impact of social factors in health outcomes.