Towards Robust and Fair Next Visit Diagnosis Prediction under Noisy Clinical Notes with Large Language Models

Clinical data quality directly impacts AI fairness: The study found that text corruption in medical notes (simulating real-world errors) not only reduces prediction accuracy but "unevenly affects demographic subgroups," meaning some patient populations experience larger performance drops than others—a critical equity concern for clinical AI systems.

Standard LLMs struggle with noisy medical information: State-of-the-art large language models showed significant reliability issues when processing degraded clinical notes, highlighting that deploying these systems in real healthcare settings requires explicit safeguards, not just reliance on general model capabilities.

Hierarchical reasoning mimics clinical best practices: The researchers developed a "hierarchical chain-of-thought strategy that emulates clinicians' reasoning," which improved both robustness and fairness. This suggests that AI systems perform better when designed to reflect how human doctors actually think through diagnostic decisions rather than treating diagnosis as a black-box prediction task.

Demographic disparities can be hidden in AI performance metrics: The paper emphasizes that overall accuracy metrics can mask "subgroup instability"—meaning an AI system might appear reliable on average while actually being unreliable for specific patient populations, underscoring the need for disaggregated fairness evaluation in clinical AI.

Proactive testing for real-world conditions is essential: By systematically testing LLMs under "diverse text corruption scenarios," the authors demonstrate that responsible AI deployment requires anticipating and testing how systems perform under realistic degraded conditions, not just ideal data.