Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Do LLMs Provide Consistent Answers to Health-Related Questions across Languages?

Ipek Baris Schlicht, Zhixue Zhao, Burcu Sayin, Lucie Flek, Paolo Rosso

TL;DR

This study addresses the risk of inconsistent health information from LLMs across languages by extending the HealthFC dataset to Turkish and Chinese and introducing a prompt-based workflow that dissects long-form answers into information units for cross-language comparison across EN, DE, TR, and ZH. Using disease-category tagging via ICD10 and multilingual translations, the authors evaluate four multilingual LLMs and find that while answer summaries are relatively stable, critical components such as Clinical Guidelines and Evidence and Health Benefits and Outcomes show pronounced cross-language inconsistencies, with notable omissions in Turkish. The results, quantified with inter-language agreement metrics (e.g., in some cases, $\,\kappa\$ values of 0.66 for TR and 0.71 for ZH, versus 0.50 for DE), highlight systemic gaps in cross-lingual alignment and the potential for misinformation to propagate differently across languages. The work underscores the need for robust cross-lingual safeguards in health AI and suggests future directions toward open models and broader multilingual coverage to improve equitable access to accurate health information across languages.

Abstract

Equitable access to reliable health information is vital for public health, but the quality of online health resources varies by language, raising concerns about inconsistencies in Large Language Models (LLMs) for healthcare. In this study, we examine the consistency of responses provided by LLMs to health-related questions across English, German, Turkish, and Chinese. We largely expand the HealthFC dataset by categorizing health-related questions by disease type and broadening its multilingual scope with Turkish and Chinese translations. We reveal significant inconsistencies in responses that could spread healthcare misinformation. Our main contributions are 1) a multilingual health-related inquiry dataset with meta-information on disease categories, and 2) a novel prompt-based evaluation workflow that enables sub-dimensional comparisons between two languages through parsing. Our findings highlight key challenges in deploying LLM-based tools in multilingual contexts and emphasize the need for improved cross-lingual alignment to ensure accurate and equitable healthcare information.

Do LLMs Provide Consistent Answers to Health-Related Questions across Languages?

TL;DR

This study addresses the risk of inconsistent health information from LLMs across languages by extending the HealthFC dataset to Turkish and Chinese and introducing a prompt-based workflow that dissects long-form answers into information units for cross-language comparison across EN, DE, TR, and ZH. Using disease-category tagging via ICD10 and multilingual translations, the authors evaluate four multilingual LLMs and find that while answer summaries are relatively stable, critical components such as Clinical Guidelines and Evidence and Health Benefits and Outcomes show pronounced cross-language inconsistencies, with notable omissions in Turkish. The results, quantified with inter-language agreement metrics (e.g., in some cases, values of 0.66 for TR and 0.71 for ZH, versus 0.50 for DE), highlight systemic gaps in cross-lingual alignment and the potential for misinformation to propagate differently across languages. The work underscores the need for robust cross-lingual safeguards in health AI and suggests future directions toward open models and broader multilingual coverage to improve equitable access to accurate health information across languages.

Abstract

Equitable access to reliable health information is vital for public health, but the quality of online health resources varies by language, raising concerns about inconsistencies in Large Language Models (LLMs) for healthcare. In this study, we examine the consistency of responses provided by LLMs to health-related questions across English, German, Turkish, and Chinese. We largely expand the HealthFC dataset by categorizing health-related questions by disease type and broadening its multilingual scope with Turkish and Chinese translations. We reveal significant inconsistencies in responses that could spread healthcare misinformation. Our main contributions are 1) a multilingual health-related inquiry dataset with meta-information on disease categories, and 2) a novel prompt-based evaluation workflow that enables sub-dimensional comparisons between two languages through parsing. Our findings highlight key challenges in deploying LLM-based tools in multilingual contexts and emphasize the need for improved cross-lingual alignment to ensure accurate and equitable healthcare information.
Paper Structure (9 sections, 5 figures, 2 tables)

This paper contains 9 sections, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methodology
  4. Information Seeker Queries
  5. Prompt-based Evaluation Workflow
  6. Experiments and Results
  7. Conclusion
  8. Acknowledgments.
  9. Disclosure of Interests.

Figures (5)

  • Figure 1: The proposed framework to automatically measure the consistency of health-related answers in EN, ZH, TR and DE.
  • Figure 2: Distribution of 11 disease categories, each with at least 20 samples.
  • Figure 3: The answers to the question "Does discontinuation of denosumab increase the risk of vertebral fractures?": the EN answer on the left and the TR answer on the right. Both were parsed by GPT-4o using the proposed discourse ontology.
  • Figure 4: The LLMs tend to produce lengthy answers to the EN and DE questions.
  • Figure 5: Answer occurence across the categories. TR answers omit some categories.