TL;DR
This paper introduces Multi-OphthaLingua, the first multilingual benchmark for ophthalmology question answering, revealing significant performance disparities in LLMs across languages, particularly for Low and Middle-Income Countries (LMICs). To address this, they propose CLARA, a novel agentic debiasing system that significantly improves accuracy and reduces bias, paving the way for more equitable AI deployment in global healthcare.
What Problem Does This Paper Address?
The healthcare domain, particularly ophthalmology, stands to benefit immensely from Large Language Models (LLMs) for tasks like triaging, clinical decision support, and report summarization. However, LLMs exhibit a well-documented performance gap across languages, often performing poorly in languages prevalent in LMICs due to less training data. This linguistic bias exacerbates existing healthcare disparities, especially in regions where specialized medical personnel are scarce. Existing debiasing methods struggle to bridge this gap effectively, particularly in specialized medical domains. This paper aims to quantify this bias in ophthalmology QA and propose a novel solution to mitigate it.
Methods
The authors introduce CLARA (Cross-Lingual Reflective Agentic system), a novel inference-time debiasing method. CLARA employs a multi-agent approach:
- Translation and Initial Evaluation Agent: Translates the input query to English and evaluates translation quality and the LLM's certainty.
- Knowledge Augmentation Agent: Utilizes a weighted Retrieval Augmented Generation (RAG) approach. It queries a vector database, assigning weights based on uncertainty. For domain-specific jargon, it expands terms using an ophthalmology dictionary before retrieval.
- Iterative Evaluation (of Retrieval) and Refinement Agent: Acts as a critic, assessing retrieved information. If insufficient, it triggers a web search. This iterative process continues until sufficient relevant information is gathered.
- Question Rewriting Agent: If RAG and web searches fail, this agent reformulates complex or nested queries to improve LLM comprehension.
This agentic framework combines pre-translation, evaluation, corrective RAG, web search, and query transformation.
Datasets
- Multi-OphthaLingua: A newly introduced multilingual ophthalmological question-answering benchmark. It comprises 1184 manually curated, parallel questions across seven languages: English, Spanish, Filipino, Portuguese, Mandarin, French, and Hindi. The questions are multiple-choice and were reviewed by ophthalmologists globally to ensure robustness and cultural appropriateness. This dataset is unique in providing paired questions across languages for direct cross-linguistic comparisons.
- RAG Corpora: The system utilizes three sources for RAG: PubMed (biomedical abstracts), medical textbooks, and Wikipedia.
Key Insights and Ideas
- LMIC Languages are Significantly Underserved: The performance gap between English and languages like Filipino and Hindi is substantial, highlighting a critical barrier to equitable AI deployment in healthcare.
- Domain-Specific Knowledge is Crucial: LLMs struggle not just with language but also with specialized medical terminology and nuances, leading to potentially critical misinterpretations.
- Existing Debiasing Methods Fall Short: Simple methods like translation or basic RAG are insufficient to close the performance gap or address the complex linguistic and domain-specific challenges in LMIC contexts.
- Cultural Nuances Matter: LLMs' responses can be influenced by cultural context, leading to inconsistent advice (e.g., emphasis on family support in Spanish vs. English post-op care).
- CLARA Offers a Holistic Solution: The multi-agent, iterative approach of CLARA, combining translation, retrieval, evaluation, and rewriting, proves effective in both improving overall accuracy and reducing cross-lingual bias.
Key Results
- Performance Disparities: GPT-4 showed a 11.6% and 12.8% accuracy drop in Filipino and Hindi, respectively, compared to English. Llama-3 70B had a 14% gap for Filipino vs. English.
- CLARA's Effectiveness:
- For GPT-4, CLARA improved overall accuracy from 63.4% (Direct Inference) to 72.2% and reduced the Filipino-English gap from 11.6% to 5.1%.
- For GPT-3.5, CLARA improved accuracy from 46.2% to 53.6% and reduced the Filipino-English gap from 11.4% to 7.9%.
- Ablation Study: Each component of CLARA (translation, web search, RAG) contributed to performance gains, with the full CLARA system achieving the best results. Notably, CLARA also improved English performance, indicating enhanced overall capability.
Limitations and Future Work
- Ethical Concerns with Less Advanced LLMs: While CLARA shows promise, the persistent limitations in less advanced models like GPT-3.5 raise ethical questions about deploying medical AI systems in real-world settings.
- Continued Research Needed: The authors emphasize the need for continued research and careful consideration of AI's role in healthcare, particularly concerning equitable access and responsible deployment.
- Future Work: The paper implicitly suggests further exploration of agentic frameworks and domain-specific knowledge integration for LLMs in healthcare. The release of the Multi-OphthaLingua benchmark is intended to foster future research in this area.
Summary updated