K-QA: A Real-World Medical Q&A Benchmark

Itay Manes, Naama Ronn, David Cohen, Ran Ilan Ber, Zehavi Horowitz-Kugler, Gabriel Stanovsky

Introduction

Large Language Models (LLMs) have demonstrated remarkable performance on medical licensing exams, such as the USMLE. However, these "textbook" benchmarks often fail to reflect the complexities of real-world clinical practice. Patients do not present their concerns in structured, multiple-choice formats; instead, they use colloquial language, describe overlapping symptoms, and seek nuanced advice. To bridge this gap, researchers from K Health and The Hebrew University of Jerusalem introduced K-QA, a benchmark designed to evaluate LLMs on authentic patient-physician interactions.

The Data Gap in Medical AI

Existing benchmarks like MedQA or BioASQ primarily focus on academic medical knowledge. While useful for testing a model's information retrieval, they suffer from two primary limitations:

  1. Data Gap: They lack the non-medical jargon and multi-faceted queries typical of real patients.
  2. Evaluation Gap: Traditional metrics like ROUGE or BLEU measure linguistic similarity rather than clinical factuality. In medicine, a high similarity score is meaningless if the model misses a critical contraindication or provides a dangerous recommendation.

K-QA addresses these issues by utilizing 1,212 de-identified questions from real users on the K Health platform, paired with a rigorous, physician-led evaluation framework.

Methodology: Expert Curation and Atomic Decomposition

The researchers developed a multi-stage pipeline to ensure the benchmark's clinical validity:

1. Expert Curation

A team of 12 physicians drafted "Gold" answers for a subset of 201 questions. These answers were based on high-quality evidence from sources like UpToDate and PubMed, ensuring a ground truth that reflects current medical standards.

2. Atomic Decomposition

To enable granular evaluation, doctors broke these gold answers down into approximately eight "minimal semantic content units" or atomic statements. Each statement was then categorized by importance:

  • Must Have: Clinically essential information (e.g., "Seek immediate care if you experience chest pain").
  • Nice to Have: Supplemental context that adds value but is not critical for safety.

3. NLI-Based Metrics

To automate the evaluation of model responses, the authors employed Natural Language Inference (NLI). Using GPT-4 as a judge, the framework measures:

  • Comprehensiveness: The percentage of "Must Have" statements from the gold answer that are entailed by the model's response.
  • Hallucination Rate: The frequency with which a model's response contradicts the gold-standard statements.

Key Results and Analysis

The study evaluated several prominent models, including GPT-4, GPT-3.5, PALM-2, and Mistral 7B, yielding several critical insights:

The "Textbook" Mirage

One of the most striking findings was that MedAlpaca, a model specifically fine-tuned on medical exam data, performed the worst on the K-QA benchmark. This suggests that proficiency in academic medical testing does not necessarily translate to being a safe or effective clinical assistant for laypeople.

The Challenge of Clinical Completeness

Even state-of-the-art models struggled with comprehensiveness. GPT-4, the top performer, captured only about 67.7% of the "Must Have" clinical information. While LLMs often produce fluent and seemingly helpful responses, they frequently omit critical safety warnings or red-flag symptoms.

The Role of Retrieval-Augmented Generation (RAG)

Implementing domain-specific RAG (using trusted sources like Mayo Clinic and the NHS) significantly improved safety. For instance, GPT-3.5 combined with RAG achieved the lowest hallucination rate (15.4%). Interestingly, the authors noted a "Model Size Paradox": larger models without RAG sometimes hallucinated more than smaller ones, likely due to overconfidence in their internal, potentially flawed training data.

The Danger of Subtle Errors

The benchmark highlighted the risk of "fluent hallucinations." In one instance, a model claimed there were "no known major interactions" between Buspar and Zoloft. In contrast, the physician's gold answer explicitly warned of the risk of serotonin syndrome. Such errors are particularly dangerous because they are presented with the same authority as correct information.

Conclusion and Future Directions

K-QA demonstrates that current LLMs, while promising, are not yet ready for autonomous clinical interaction. The automated NLI evaluation showed an 83% agreement rate with human physicians, validating it as a scalable proxy for manual audits.

Future work will focus on expanding the dataset and refining the scoring system to penalize critical clinical errors—such as missing a life-threatening warning—more heavily than minor omissions. As medical AI continues to evolve, benchmarks like K-QA will be essential for ensuring that models are not just knowledgeable, but safe and reliable in real-world settings.

Summary updated

BibTeX
@inproceedings{manes2024k,
  author = {Itay Manes and Naama Ronn and David Cohen and Ran Ilan Ber and Zehavi Horowitz-Kugler and Gabriel Stanovsky},
  title = {{K-QA: A Real-World Medical Q\&A Benchmark}},
  booktitle = {BioNLP @ ACL},
  year = {2024},
  doi = {10.18653/v1/2024.bionlp-1.22},
}
Figure