Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Mitigating hallucinations in large language models for medical question answering: a systematic review
ABSTRACT
Background:
Large language models deployed in medical question-answering systems generate hallucinations that pose direct patient safety risks, with error rates ranging from near-zero in constrained tasks to 24.6% in complex reasoning scenarios.
Objective:
This systematic review aimed to characterise hallucination types distinctive to medical question-answering systems and to evaluate the effectiveness and limitations of current mitigation strategies under clinical constraints.
Methods:
We conducted a systematic review following PRISMA 2020 guidelines, registered prospectively in PROSPERO (CRD420251024791). Eight bibliographic databases were searched for articles from January 2019 to December 2025. Three reviewers screened records using a Population-Concept-Context eligibility framework; pre-production calibration across all three reviewers yielded Cohen's kappa of 0.83, and full-text eligibility was assessed by two independent reviewers per record (κ = 0.78). Evidence was synthesised narratively using the SWiM framework, stratified by hallucination type and mitigation technique family.
Results:
A total of 177 studies were included. Hallucinations manifest across two primary dimensions: factuality errors (incorrect medical knowledge, fabricated information) and faithfulness errors (instruction misalignment, logical inconsistency). Retrieval-augmented generation demonstrated the most consistent benefit across included studies, corroborated by an external meta-analysis reporting a pooled odds ratio of 1.35 (95% CI: 1.19–1.53). Domain-specific fine-tuning paradoxically increased hallucination in some contexts, with biomedical models underperforming general-purpose counterparts by up to 34% points on unseen clinical tasks. Benchmark performance frequently diverged from real-world clinical reliability.
Conclusions:
Effective deployment of medical question-answering systems for real-world applications requires dual-strategy approaches combining retrieval-augmented generation with structured prompting, mandatory detection mechanisms for multimodal systems, and evaluation frameworks prioritising clinical utility over examination performance. Current approaches remain predominantly reactive; the field must prioritise architectural changes to enable proactive hallucination prevention. Clinical Trial: PROSPERO CRD420251024791; https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=1024791
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.