Bias Amplification in Intersectional Subpopulations for Clinical Phenotyping by Large Language Models

Abstract

Abstract Large Language Models (LLMs) have demonstrated remarkable performance across diverse clinical tasks. However, there is growing concern that LLMs may amplify human bias and reduce performance quality for vulnerable subpopulations. Therefore, it is critical to investigate algorithmic underdiagnosis in clinical notes, which represent a key source of information for disease diagnosis and treatment. This study examines prevalence of bias in two datasets - smoking and obesity - for clinical phenotyping. Our results demonstrate that state-of-the-art language models selectively and consistently underdiagnosed vulnerable intersectional subpopulations such as young-aged-males for smoking and middle-aged-females for obesity. Deployment of LLMs with such biases risks skewing clinicians’ decision-making which may lead to inequitable access to healthcare. These findings emphasize the need for careful evaluation of LLMs in clinical practice and highlight the potential ethical implications of deploying such systems in disease diagnosis and prognosis.

Ähnliche Arbeiten