[Paper Review] MediQ: Question-Asking LLMs and a Benchmark for Reliable Interactive Clinical Reasoning

Shuyue Stella Li, Vidhisha Balachandran+5 more|arXiv (Cornell University)|Jun 3, 2024

Artificial Intelligence in Law4 citations

TL;DR

This paper introduces MediQ, a framework for interactive, question-asking LLMs in clinical reasoning that simulates realistic patient-doctor dialogues with a Patient system and an adaptive Expert system. It shows that state-of-the-art LLMs like GPT-3.5 and Llama-3 struggle with proactive information seeking, but performance improves by 22.3% with a novel abstention module for confidence estimation, though still lags behind full-information baselines.

ABSTRACT

Users typically engage with LLMs interactively, yet most existing benchmarks evaluate them in a static, single-turn format, posing reliability concerns in interactive scenarios. We identify a key obstacle towards reliability: LLMs are trained to answer any question, even with incomplete context or insufficient knowledge. In this paper, we propose to change the static paradigm to an interactive one, develop systems that proactively ask questions to gather more information and respond reliably, and introduce an benchmark - MediQ - to evaluate question-asking ability in LLMs. MediQ simulates clinical interactions consisting of a Patient System and an adaptive Expert System; with potentially incomplete initial information, the Expert refrains from making diagnostic decisions when unconfident, and instead elicits missing details via follow-up questions. We provide a pipeline to convert single-turn medical benchmarks into an interactive format. Our results show that directly prompting state-of-the-art LLMs to ask questions degrades performance, indicating that adapting LLMs to proactive information-seeking settings is nontrivial. We experiment with abstention strategies to better estimate model confidence and decide when to ask questions, improving diagnostic accuracy by 22.3%; however, performance still lags compared to an (unrealistic in practice) upper bound with complete information upfront. Further analyses show improved interactive performance with filtering irrelevant contexts and reformatting conversations. Overall, we introduce a novel problem towards LLM reliability, an interactive MediQ benchmark and a novel question-asking system, and highlight directions to extend LLMs' information-seeking abilities in critical domains.

Motivation & Objective

• To address the unreliability of LLMs in clinical reasoning when given incomplete information.
• To simulate realistic medical consultations where patients provide incomplete data and clinicians must proactively seek missing details.
• To develop a benchmark that evaluates LLMs in interactive, multi-turn clinical reasoning rather than static, single-turn QA.
• To investigate whether and how LLMs can be adapted to ask follow-up questions instead of guessing when uncertain.
• To identify key challenges and performance gaps in current LLMs' ability to engage in reliable, adaptive information-seeking during medical diagnosis.

Proposed method

• MediQ framework consists of a Patient system that simulates a patient and answers follow-up questions using factual recall from patient records.
• An Expert system acts as a doctor’s assistant, deciding at each turn whether to answer or ask for more information based on confidence.
• The framework converts MedQA and Craft-MD into interactive setups by initially providing only partial patient information.
• A novel abstention module is introduced to improve confidence estimation and decision-making on whether to ask more questions.
• Multiple prompt variations (system and response prompts) are tested to evaluate robustness and consistency across model configurations.
• Performance is evaluated using diagnostic accuracy across medical specialties and question difficulty levels, with comparisons to full-information upper bounds.

Figure 1 : Information Seeking Task. In standard medical QA tasks (left), all necessary information is given to the assistant model at the same time. When given partial information, current LLMs only provides general responses (middle). In a more realistic scenario (right), the presentation of patie

Experimental results

Research questions

• RQ1 Can state-of-the-art LLMs reliably engage in interactive clinical reasoning by asking follow-up questions when information is incomplete?
• RQ2 How does the performance of LLMs in interactive clinical reasoning compare to their performance in single-turn QA with full information?
• RQ3 What impact do different prompt engineering strategies (system and response prompts) have on the reliability and consistency of LLMs in information-seeking tasks?
• RQ4 How does the difficulty of medical questions and specialty type affect the benefit of interactive information seeking?
• RQ5 What components, such as confidence estimation or context filtering, most significantly improve the diagnostic accuracy of interactive LLMs?

Key findings

• Direct prompting of SOTA LLMs to ask questions results in a relative 11.3% drop in diagnostic accuracy compared to single-turn inference with the same limited context.
• The introduction of a confidence-based abstention module improves diagnostic accuracy by 22.3% across datasets, demonstrating the value of deliberate uncertainty handling.
• Despite improvements, the best interactive Expert system still lags by 10.3% behind the upper bound when full information is provided upfront.
• Performance gains are most pronounced in challenging specialties like ophthalmology, where accuracy increased from 18.2% to 45.5% with interaction.
• Prompt variations such as rationale generation modestly improve performance, while answer shuffling has no significant effect, indicating robustness to certain biases.
• The study reveals that LLMs tend to rely on parametric knowledge when context is sparse, leading to high generality agreement (50.0%) with the most common option, which undermines accuracy.

Figure 2 : The MediQ Framework. MediQ operationalizes a realistic dynamic medical consultation process. It consists of a Patient system and an Expert system to evaluate medical info-seeking.