Boosting Social Determinants of Health Extraction with Semantic Knowledge Augmented Large Language Model.

Area of Science:

• Health Informatics
• Natural Language Processing
• Clinical Data Mining

Background:

• Social determinants of health (SDoH) significantly impact health outcomes and contribute to health disparities.
• Extracting SDoH from unstructured Electronic Health Records (EHRs) is challenging due to data scarcity and imbalanced categories.
• Large Language Models (LLMs) show promise for SDoH extraction but struggle with data imbalance.

Purpose of the Study:

• To improve automated extraction of SDoH information from clinical narratives.
• To address the performance limitations of LLMs caused by imbalanced SDoH data.
• To enhance LLM feature representations for underrepresented SDoH classes.

Main Methods:

• Augmenting LLMs with semantic knowledge from the Unified Medical Language Systems (UMLS).
• Implementing a data augmentation strategy to generate semantically enriched clinical narratives during LLM pre-finetuning.
• Utilizing publicly available MIMIC-SDoH data for extensive experimentation.

Main Results:

• The proposed approach significantly improves SDoH extraction accuracy.
• Enhanced performance is particularly notable for imbalanced SDoH categories.
• Semantic enrichment during pre-finetuning leads to better LLM adaptation and initialization.

Conclusions:

• Augmenting LLMs with UMLS semantic knowledge is an effective strategy for improving SDoH extraction from EHRs.
• The data augmentation method enhances LLM performance on imbalanced datasets, crucial for addressing health disparities.
• This approach offers a promising solution for more accurate and equitable SDoH data analysis in healthcare.