Spurious Correlations and Beyond: Understanding and Mitigating Shortcut Learning in SDOH Extraction with Large Language Models

TL;DR

This paper investigates how large language models for social determinants of health extraction can rely on superficial cues, leading to false predictions and disparities, and explores mitigation strategies to improve their reliability in clinical text analysis.

Contribution

It identifies spurious correlations in LLM-based SDOH extraction and evaluates mitigation techniques like prompt engineering and reasoning chains to enhance model robustness.

Findings

Mentions of alcohol or smoking cause false drug use predictions.

Gender disparities affect model performance.

Mitigation strategies reduce false positives.

Abstract

Social determinants of health (SDOH) extraction from clinical text is critical for downstream healthcare analytics. Although large language models (LLMs) have shown promise, they may rely on superficial cues leading to spurious predictions. Using the MIMIC portion of the SHAC (Social History Annotation Corpus) dataset and focusing on drug status extraction as a case study, we demonstrate that mentions of alcohol or smoking can falsely induce models to predict current/past drug use where none is present, while also uncovering concerning gender disparities in model performance. We further evaluate mitigation strategies - such as prompt engineering and chain-of-thought reasoning - to reduce these false positives, providing insights into enhancing LLM reliability in health domains.