Metamorphic Testing for Robustness and Fairness Evaluation of LLM-based Automated ICD Coding Applications

Abstract

Healthcare and medical domain-specific LLMs (BioMed LLMs), such as PubMedBERT and Med-PaLM, are developed and pre-trained on biomedical and clinical text to be used specifically in healthcare and medical applications. The recent popularity of these BioMed LLMs increased the use of LLMs in health and medical applications to perform various critical tasks, including ICD (International Classification of Diseases) coding. For such safety-critical applications, it is vital to focus not just on accuracy but also on other quality attributes such as robustness and fairness. Unfortunately, application developers rarely assess these attributes despite their importance in BioMed LLMs-based applications due to difficulties in defining the expected output. This study uses Metamorphic Testing (MT) to evaluate the robustness and fairness of the BioMed LLM-based automated ICD coding application. We defined several Metamorphic Relations (MRs) to evaluate these quality attributes systematically. Our results using the MIMIC-III dataset reveal several instances where the application performance is significantly impacted due to various simple manipulations that mimic common mistakes in the input clinical notes. Our findings highlight the necessity of rigorous testing for these metrics to ensure the reliable use of BioMed LLMs in healthcare and medical applications. Further, our research provides a comprehensive framework for such evaluations by leveraging MT, which is helpful to the application developers and contributes to developing more reliable and robust biomedical AI systems.