Structured Insight from Unstructured Data: Large Language Models for SDOH-Driven Diabetes Risk Prediction.

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Social Determinants of Health Research

Background:

• Social determinants of health (SDOH) significantly impact Type 2 Diabetes (T2D) management but are often missing from clinical data.
• Current structured screening tools for SDOH lack the depth to capture patient complexity.

Purpose of the Study:

• To explore using large language models (LLMs) to extract structured SDOH data from unstructured patient narratives.
• To evaluate the predictive power of LLM-extracted SDOH features and patient narratives for diabetes control.
• To integrate narrative data into conventional risk prediction models.

Main Methods:

• Collected unstructured life story interviews from 65 T2D patients (aged 65+).
• Utilized LLMs with retrieval-augmented generation for qualitative summarization and quantitative SDOH rating.
• Applied structured SDOH ratings and lab biomarkers to machine learning models (Ridge, Lasso, Random Forest, XGBoost).
• Assessed LLM accuracy in predicting diabetes control levels directly from interview text (A1C redacted).

Main Results:

• LLMs successfully extracted structured SDOH information from patient narratives.
• LLM-derived SDOH ratings and narrative data showed predictive value for diabetes control.
• LLMs achieved 60% accuracy in predicting diabetes control levels from interview text alone.

Conclusions:

• LLMs offer a scalable method to convert unstructured SDOH data into actionable clinical insights.
• Integrating LLM-processed narrative data can augment existing clinical risk prediction models.
• This approach enhances understanding of patient experiences and social contexts in T2D management.