MedExQA: Medical Question Answering Benchmark with Multiple Explanations

TL;DR

MedExQA tackles the gap in medical QA benchmarks by introducing two explanations per QA pair across five underrepresented specialties, enabling evaluation of model reasoning and explainability beyond accuracy. The authors benchmark 18 open-source baselines and GPT models, and introduce MedPhi-2, a Phi-2–based medical model pretrained on domain data and instruction-tuned, which outperforms many baselines in explanation generation. They employ a multi-faceted evaluation including logits-based MCQ accuracy, chat-style responses, automated explanation metrics, and human judgments, finding that generation-based evaluation with multiple explanations aligns more closely with human assessment. The benchmark reveals strong performance of instruction-tuned medical models and persistent challenges in Speech Language Pathology, underscoring the need for targeted domain pretraining and rigorous explainability criteria with open data release to accelerate progress in medical LLMs.

Abstract

This paper introduces MedExQA, a novel benchmark in medical question-answering, to evaluate large language models' (LLMs) understanding of medical knowledge through explanations. By constructing datasets across five distinct medical specialties that are underrepresented in current datasets and further incorporating multiple explanations for each question-answer pair, we address a major gap in current medical QA benchmarks which is the absence of comprehensive assessments of LLMs' ability to generate nuanced medical explanations. Our work highlights the importance of explainability in medical LLMs, proposes an effective methodology for evaluating models beyond classification accuracy, and sheds light on one specific domain, speech language pathology, where current LLMs including GPT4 lack good understanding. Our results show generation evaluation with multiple explanations aligns better with human assessment, highlighting an opportunity for a more robust automated comprehension assessment for LLMs. To diversify open-source medical LLMs (currently mostly based on Llama2), this work also proposes a new medical model, MedPhi-2, based on Phi-2 (2.7B). The model outperformed medical LLMs based on Llama2-70B in generating explanations, showing its effectiveness in the resource-constrained medical domain. We will share our benchmark datasets and the trained model.

Figures (6)

• Figure 1: 2D t-SNE plot for MedExQA, MedQA, MedMCQA, and MMLU (Medicine Related 9 subjects) datasets.
• Figure 2: Scatter plot of model performance. The Y-axis is the macro average of accuracy based on logits (Table \ref{['tab:result_choice']}). The X-axis is the average score of generated explanations (Table \ref{['tab:result_exp']}). The dot size is proportional to the model size.
• Figure 3: Human evaluation on the generated explanations, which scales from 0 to 5. The models in the legend are ordered by macro average from lowest to highest. Only models passed (3 or above) in at least one of the specialties are included.
• Figure 4: Example of data in Speech Language Pathology and Qualitative Analysis Example. Two sets of explanations, Explanation 1 and Explanation 2, are provided. The score given by humans is provided beneath the model name. The response with no explanations has a grey font color. Red shows the irrelevant or wrong sentences or phrases. Yellow demonstrates incoherent phrases or errors. Green highlights coherent and correct sentences.
• Figure 5: Word Count Distribution Plots for Explanations. Top: Explanation 1. Bottom: Explanation 2.