BanglaDiaHallu: Hallucinations of large language models and the effect of guideline-grounded retrieval-augmented generation in Bangla diabetes question answering
- 1Jahangirnagar University
Abstract
Large language models (LLMs) are increasingly consulted for health advice, but they can produce fluent statements that are false or unsupported, known as hallucinations, and this risk is poorly characterised in low-resource languages such as Bangla. We measured hallucinations of locally deployable open-weight LLMs in Bangla diabetes question answering and tested whether retrieval-augmented generation (RAG) grounded in Bangladeshs National Guideline on Diabetes Mellitus reduces them. Forty patient questions spanning six clinical sub-domains were each answered twice, under zero-shot prompting, by Qwen3.5-9B, TigerLLM-9B, Qwen3.6-35B-A3B and Gemma-4-26B-A4B, and by BanglaDiaRAG, which adds hybrid dense (BGE-M3) and sparse (BM25) retrieval from a physician-reviewed Bangla guideline knowledge base to Qwen3.5-9B. Two physicians rated the 400 de-identified responses by consensus for medical accuracy, completeness, safety, readability and reliability on a 0-1 scale, and responses with accuracy below 0.70 were classified as hallucinated. Hallucination rates were 2.5% for Gemma-4-26B-A4B, 3.8% for TigerLLM-9B, 22.5% for Qwen3.6-35B-A3B and 77.5% for Qwen3.5-9B. Guideline grounding lowered the rate of Qwen3.5-9B to 11.3% (difference -66.3 percentage points; 95% CI -76.3 to -55.0) and raised every criterion by 0.24 to 0.38 points (Holm-adjusted p < 0.001). BanglaDiaRAG outscored Qwen3.6-35B-A3B on all five criteria and matched or exceeded TigerLLM-9B, but it remained below Gemma-4-26B-A4B in safety, readability and reliability, and its residual errors were specific, confident and potentially harmful. Guideline grounding can turn a weak general-purpose model into a much safer Bangla diabetes assistant, but the choice of base model mattered as much as retrieval and no configuration was free of hallucination, so clinical oversight remains essential before patient-facing use.
Related
- Development and evaluation of Retrieval-Augmented Generation (RAG) and GraphRAG for guideline-based clinical decision support in chronic kidney disease
- Retrieval-augmented generation in biomedicine: a survey of technologies, datasets, and clinical applications
- Retrieval-augmented generation in medicine: A scoping review of technical implementations, clinical applications, and ethical considerations
- IPF-RAG: temporal-aware retrieval-augmented generation for idiopathic pulmonary fibrosis care.
- Limits of Single-Pass Retrieval-Augmented Generation for AI-Powered Cancer Care Navigation: A Comparison of Retrieval Strategies
BibTeX
@misc{nahid2026bangladiahallu,
title = {{BanglaDiaHallu: Hallucinations of large language models and the effect of guideline-grounded retrieval-augmented generation in Bangla diabetes question answering}},
author = {K. Nahid and M. Anam and M. T. H. Saon and O. H. Any and K. M. R. Hassan and M. M. Anwar},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.14.26365587},
url = {https://www.medrxiv.org/content/10.64898/2026.10.14.26365587v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - BanglaDiaHallu: Hallucinations of large language models and the effect of guideline-grounded retrieval-augmented generation in Bangla diabetes question answering
AU - K. Nahid
AU - M. Anam
AU - M. T. H. Saon
AU - O. H. Any
AU - K. M. R. Hassan
AU - M. M. Anwar
PY - 2026
DA - 2026/10/14
PB - medRxiv
DO - 10.64898/2026.10.14.26365587
UR - https://www.medrxiv.org/content/10.64898/2026.10.14.26365587v1
AB - Large language models (LLMs) are increasingly consulted for health advice, but they can produce fluent statements that are false or unsupported, known as hallucinations, and this risk is poorly characterised in low-resource languages such as Bangla. We measured hallucinations of locally deployable open-weight LLMs in Bangla diabetes question answering and tested whether retrieval-augmented generation (RAG) grounded in Bangladeshs National Guideline on Diabetes Mellitus reduces them. Forty patient questions spanning six clinical sub-domains were each answered twice, under zero-shot prompting, by Qwen3.5-9B, TigerLLM-9B, Qwen3.6-35B-A3B and Gemma-4-26B-A4B, and by BanglaDiaRAG, which adds hybrid dense (BGE-M3) and sparse (BM25) retrieval from a physician-reviewed Bangla guideline knowledge base to Qwen3.5-9B. Two physicians rated the 400 de-identified responses by consensus for medical accuracy, completeness, safety, readability and reliability on a 0-1 scale, and responses with accuracy below 0.70 were classified as hallucinated. Hallucination rates were 2.5% for Gemma-4-26B-A4B, 3.8% for TigerLLM-9B, 22.5% for Qwen3.6-35B-A3B and 77.5% for Qwen3.5-9B. Guideline grounding lowered the rate of Qwen3.5-9B to 11.3% (difference -66.3 percentage points; 95% CI -76.3 to -55.0) and raised every criterion by 0.24 to 0.38 points (Holm-adjusted p < 0.001). BanglaDiaRAG outscored Qwen3.6-35B-A3B on all five criteria and matched or exceeded TigerLLM-9B, but it remained below Gemma-4-26B-A4B in safety, readability and reliability, and its residual errors were specific, confident and potentially harmful. Guideline grounding can turn a weak general-purpose model into a much safer Bangla diabetes assistant, but the choice of base model mattered as much as retrieval and no configuration was free of hallucination, so clinical oversight remains essential before patient-facing use.
ER -