AI xPress

AI and computer science research.

Back to Foragentis

Healthcare & Bio AI

Fine-tuned language models identified 93.8% of patients with adverse social determinants of health in clinical notes, while ICD-10 diagnosis codes captured 2.0%

Healthcare & Bio AI November 20, 2026

A pharmacy counting tray with tablets

In a 2024 paper in npj Digital Medicine, researchers report that large language models can extract social determinants of health from the narrative text of electronic health records. Social determinants are life conditions such as employment, housing, transportation, parental status, relationship, and social support. The best fine-tuned models identified 93.8% of patients with adverse social determinants of health, compared with 2.0% captured by ICD-10 diagnosis codes.

What did the researchers test?

Researchers tested models from the Flan-T5 family and ChatGPT-family models for finding six social determinant categories in electronic health record notes. The best fine-tuned model for any social determinant mention was Flan-T5 XL, with a macro-F1 score of 0.71. The best for adverse mentions was Flan-T5 XXL, with a score of 0.70. Macro-F1 is a combined measure of precision and recall across categories. Adding synthetic training data written by a large language model helped the smaller Flan-T5 models, improving their F1 scores by 0.12 to 0.23.

How did the fine-tuned models compare with ChatGPT?

The fine-tuned models outperformed ChatGPT-family models in zero-shot and few-shot settings, where the model receives no task examples or only a few examples. The exception was GPT4 with 10 examples for adverse social determinants, which did better. Fine-tuned models were also less likely to change their predictions when race or ethnicity and gender descriptors were added to the text, a result the authors describe as suggesting less algorithmic bias, with a p value below 0.05.

What are social determinants of health?

They are the conditions in which people live and work, such as employment, housing, transportation, parental status, relationship, and social support. These details are often missing or incomplete in the structured parts of a patient's record, so the study looked for them in narrative clinical notes.

Why does the result matter?

Adverse social determinants can affect patient outcomes, but they are sparsely documented. Language models could help identify patients who may benefit from resource support. The best models identified 93.8% of patients with adverse social determinants, while ICD-10 codes captured 2.0%.

Limits and what's next

The abstract describes a study of narrative text in electronic health records, but it does not report the number of records or patients, the dates, or which hospitals contributed the notes. It also notes class imbalance and data limitations as challenges. The performance figures come from this single study, and the abstract reports no independent replication.

Source: Marco Guevara et al., “Large language models to identify social determinants of health in electronic health records,” npj Digital Medicine, 2024.