PaperlayerView exact revision on arXiv

Source-linked AI summary

MediQ: Question-Asking LLMs and a Benchmark for Reliable Interactive Clinical Reasoning

Shuyue Stella Li, Vidhisha Balachandran, Shangbin Feng, Jonathan S. Ilgen, Emma Pierson, Pang Wei Koh, Yulia Tsvetkov

arXiv:2406.00922v3cs.CLcs.AI

TL;DR

Existing medical QA benchmarks usually provide complete information in a static single turn, despite clinical interactions often beginning with incomplete context. The paper introduces MEDIQ, an interactive Patient–Expert benchmark and question-asking system built around abstention and follow-up elicitation. SOTA LLMs struggle with proactive information seeking, while MEDIQ-Expert partially closes the incomplete-information gap; the benchmark remains limited by available datasets, paid Patient APIs, and multiple-choice evaluation.

  • Problem

    Existing static single-turn medical QA benchmarks do not evaluate reliable information seeking when clinical context is incomplete.

  • Method

    MEDIQ converts medical QA datasets into dynamic Patient–Expert interactions, while MEDIQ-Expert uses modular abstention and follow-up questioning before decisions.

  • Results

    SOTA LLMs struggle with proactive information seeking, while MEDIQ-Expert’s abstaining capabilities partially close the gap between incomplete- and full-information settings.

  • Takeaways & Limitations

    MEDIQ provides an experimental framework for evaluating and developing reliable LLM information seeking in clinical reasoning.

  • Takeaways & Limitations

    MEDIQ is limited by scarce detailed diagnostic datasets, a paid Patient-system API, and evaluation in multiple-choice format.

Abstract

from arXiv · show

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.

1 Introduction

MEDIQ reframes medical QA as interactive information seeking under incomplete context, with an Expert system deciding whether to answer or ask follow-up questions. The benchmark and MEDIQ-Expert expose current LLMs’ difficulty with proactive information seeking while partially improving the incomplete-information gap.

  • Framework motivation: MEDIQ replaces static single-turn medical QA with dynamic interactions between a Patient system and an Expert system.The Patient responds to follow-up questions, while the Expert gathers information before making a medical decision.
  • Benchmark construction: The practical conversion pipeline parses patient records so only partial information is initially provided, enabling interactive versions of MEDQA and CRAFT-MD.The Patient system is developed to retrieve correct facts from the full patient record when answering Expert inquiries.
  • Empirical gap: SOTA LLMs including Llama-3-Instruct, GPT-3.5, and GPT-4 struggle with proactive information seeking under incomplete initial information.They perform relatively well with complete information but struggle in more realistic interactive settings.
  • Proposed system: MEDIQ-Expert uses abstention to reduce unconfident answers and partially close the gap between incomplete- and full-information settings.The system is designed to refrain from answering when confidence is insufficient and instead elicit missing information.

2 MEDIQ: Dynamic Medical Consultation Framework Overview

MEDIQ models consultation as iterative reasoning over a partial patient record, where the Expert alternates between confidence assessment, question asking, information integration, and decision making. Its modular MEDIQ-Expert architecture operationalizes this process with abstention strategies that determine whether to ask or answer.

  • Task definition: The task begins with limited patient information, while the Patient system retains the full record needed to answer the medical question.The Expert receives the initial presentation and progressively gathers additional information through interaction.
  • Expert system breakdown: MEDIQ-Expert decomposes clinical reasoning into initial assessment, abstention, question generation, information integration, and decision making.Each step is modular and easily modifiable.
  • Expert system breakdown: The abstention module outputs whether to proceed to a final answer; otherwise, the Expert continues to question generation.It evaluates confidence using the multiple-choice question, initial presentation, and conversation log.
  • Expert system breakdown: The question generation module crafts atomic questions to elicit additional medical evidence, such as lifestyle factors and physical examination results.The information integration module appends each question-answer pair to the conversation log for subsequent assessment.
  • Abstention strategies: The BASIC baseline asks the model to generate either an atomic question or an answer, while Numerical, Binary, and Scale variants explicitly probe confidence.Numerical uses a confidence score and threshold; Binary classifies sufficiency, while Scale adds discrete confidence granularity.

3 Experiments

The experiments validate MEDIQ’s Patient system, examine how information availability relates to accuracy, and evaluate Expert information seeking across interactive and non-interactive settings. They use converted MEDQA and CRAFT-MD data, reliability metrics, and baselines with different information levels and question-asking strategies.

  • Evaluation setup: MEDQA and CRAFT-MD are converted into interactive evaluation datasets for the experiments.MEDQA contributes 10,178/1,272/1,273 train/dev/test samples, while CRAFT-MD contains 140 dermatology patient records.
  • Patient evaluation: Patient responses are evaluated for factuality and relevance, with automated metrics validated by manual annotations.Factuality measures consistency with the patient record, while relevance measures whether the response answers the Expert’s question.
  • Patient evaluation: Factuality is computed as the percentage of atomic response statements supported by the patient record, averaged across patients.Responses are decomposed into atomic statements before support is assessed.
  • Expert evaluation: Non-interactive baselines compare Full, Initial, and None information availability levels, with the question and options supplied in every setting.Full corresponds to standard QA; Initial provides gender, age, and chief complaint.
  • Expert evaluation: Vanilla LLMs answer with incomplete information rather than requesting evidence unless explicitly prompted with the option to ask follow-up questions.The BASIC question-asking baseline is introduced to study this behavior, alongside abstention variants.
  • Expert evaluation: Expert performance is evaluated using interaction efficiency, measured by number of questions, and solution accuracy.These metrics serve as proxies for the quality of clinical reasoning and information-seeking questions.

4 Results

The Patient system’s factuality varies by response setting, while Expert systems struggle to seek missing information interactively. Abstention strategies improve accuracy and questioning, but remain below the complete-information upper bound.

  • Patient system reliability: The Fact-Select Patient setting was used in subsequent experiments after Direct and Instruct variants showed factuality problems.Direct responses sometimes returned “Yes” or “No,” while Instruct responses sometimes inferred rather than recited record facts.
  • Limited information: Performance drops markedly as patient information decreases from Full to Initial to None availability levels.In BASIC, accuracy was 11.3% lower on average than the non-interactive Initial setting with the same starting information.
  • Interactive questioning: Most models ask no follow-up questions in BASIC, revealing weak proactive identification and elicitation of missing information.Within Llama and GPT families, model size correlates with questions asked and accuracy.
  • Abstention strategies: 22.3% improvement over BASIC was achieved by Scale Abstention with rationale generation and self-consistency factor 3.This strategy also outperformed non-interactive Initial by 12.1%.
  • Model scale: Models larger than 70B surpass non-interactive Initial with the best abstention, whereas smaller models still struggle.The pattern generalizes across different LLMs.
  • Remaining gap: The best MEDIQ Expert closes only 51.2% of the gap between Non-Interactive Initial and Full information scenarios.Complete information remains an upper bound that interactive systems do not reach.

5 Analysis

Analysis attributes BASIC’s performance loss to irrelevant context and conversation format, while abstention and rationale generation improve interactive reasoning by supporting confidence estimation and better questions.

  • BASIC failure modes: 11.3% relative accuracy is lost from BASIC to non-interactive Initial across benchmarked LLMs.For GPT-3.5 on iMEDQA, the relative drop is 7.43%.
  • Irrelevant context: Filtering unanswerable and repeated question-response pairs tests whether ignoring irrelevant context affects performance.The Relevant filter retains record-based questions and responses.
  • Conversation format: Converting conversations into paragraph format removes Expert questions and rewrites unanswerable questions as statements.This controls for the mismatch between dialogue format and typical document-like pretraining data.
  • Error analysis: Unique-Para surpasses BASIC by 5.7pp and NI-Initial by 2.3pp.Relevant and Unique each improve performance by 2pp, while their combined filtering effect is indistinguishable from either filter alone.
  • Abstention thresholds: Longer interactions improve accuracy, but performance still varies by abstention strategy.This suggests abstention strategies differ in confidence-estimation quality.
  • Rationale generation: Rationale generation lowers expected calibration error from 2.9 to 2.1 and helps Experts ask better questions.It produces more conservative confidence estimates, while including abstention responses in question-generation context significantly improves performance.

6 Related Work

Prior work covers medical question answering with rule-based systems and LLM agents, alongside interactive conversational models where iterative information gathering is important.

  • Medical question answering: Medical QA has progressed from rule-based systems to LLM-powered agents and includes benchmarks such as MultiMedQA.MultiMedQA contains multiple-choice and open-ended questions from varied sources.
  • Interactive models and agents: Interactive conversational models extend standard QA through dialogue in domains including customer support and negotiation.Prior studies also elicit richer human-preference information through iterative interactions.

7 Conclusion

The paper identifies a gap in LLMs’ proactive information seeking and introduces MEDIQ and MEDIQ-Expert to evaluate and improve interactive clinical reasoning. SOTA models still struggle to gather necessary information, motivating broader future extensions.

  • Conclusion: MEDIQ evaluates LLM question-asking through realistic clinical interactions with partial information initially available.The benchmark supports development of reliable models for personalized, precise, and reliable decision settings.
  • Conclusion: MEDIQ-Expert improves clinical reasoning through better confidence judgment and medical expertise.The system operationalizes interactive and explicit clinical reasoning with greater interpretability in reasoning and decision-making flows.
  • Conclusion: Llama-3 and GPT-4 struggle to gather information necessary for accurate medical decisions.The authors encourage broader Patient systems, medical knowledge integration, and interaction customization.

Limitations

MEDIQ is constrained by limited data availability, reliance on a paid Patient-system API, and evaluation primarily in multiple-choice settings.

  • Detailed patient-information datasets sufficient for diagnosis are scarce; only MEDQA and CRAFT-MD met the authors’ needs.
  • The benchmark’s Patient system relies on a paid API, leaving open the need for an open-source alternative.
  • The evaluation remains limited to multiple-choice questions, although MEDIQ can be extended to open-ended settings with suitable datasets and metrics.

Ethics Statement

The paper emphasizes that MEDIQ is an experimental prototype rather than a deployable clinical system, with privacy, bias, and over-reliance risks requiring attention.

  • MEDIQ is an initial prototype for testing LLM clinical reasoning, not a system intended for direct interaction with users.
  • Using closed-source LLMs with medical data risks confidential-information leakage and patient-privacy violations.
  • The framework may contain social and cultural biases from its LLMs, datasets, prompts, and abstention procedures.
  • Real-world use would require training users and clinicians to avoid over-reliance on error-prone technology and understand privacy and fairness risks.
  • The Patient system is designed to answer Expert follow-up questions from the full patient record using factual, context-grounded responses without inference.
  • Patient-system evaluation measures factuality and relevance, including whether responses are factually consistent with the patient context and answer the Expert’s question.

A.3 Qualitative Patient System Failure Analysis

The MEDIQ Patient System analysis identifies factuality, completeness, and role-play failures that can undermine realistic clinical interaction and subsequent Expert evaluation. Manual preference-based annotation is used to compare Patient variants on factual correctness and information omission.

  • Direct Patient: Direct Patient responses often reduce medically relevant answers to binary yes/no statements, omitting symptom detail needed for diagnosis.Examples include answering “Yes” despite records specifying subsided nausea or two prior episodes.
  • Direct Patient: Direct Patient hallucinations can introduce unsupported medical-history facts, such as claiming insulin therapy when the record lists no medications.The authors characterize these false facts as problematic because they can lead to inappropriate medical decisions.
  • Instruct Patient: The Instruct Patient variant sometimes generates provider-directed advice instead of answering from the patient’s perspective.This indicates role confusion when the conversation log is compiled.
  • Manual Evaluation: Manual evaluation compares Patient responses for factual correctness and information completeness using pairwise preferences across sampled interactions.Annotators compare responses to the same record and Expert question, while ties are excluded from win-rate calculation.
  • Expert Abstention: The Expert prompt instructs the model to answer when confident and otherwise ask one specific atomic question targeting missing information.Scale abstention defines five confidence levels, from conclusive support to near-random guessing, to guide whether further questions are needed.

C Model Version and Compute

The experiments use quantized Llama and specified GPT model versions, with confidence thresholds tuned by grid search. Binomial standard deviation is used to estimate uncertainty when repeated runs are expensive, and a sanity check evaluates this approximation.

  • Model Specifications: Llama models use 4-bit quantization for 70B variants and 8-bit quantization for 7B, 8B, and 13B variants to reduce GPU usage.GPT-3.5 and GPT-4 use the gpt-3.5-turbo-0125 and gpt-4-turbo-2024-04-09 versions, respectively.
  • Hyperparameters: Confidence thresholds determining when the Expert stops asking questions are selected through grid search rather than set arbitrarily.The threshold is treated as an important hyperparameter for both Patient and Expert systems.
  • Compute: Experiment duration ranges from roughly 30 minutes for the simplest GPT-3.5 setup to 7 days for the most demanding configurations.Runtime scales with conversation length, rationale generation, and self-consistency.
  • Statistical Testing: The study approximates confidence intervals with a binomial distribution using model accuracy p and dataset size n.The corresponding standard deviation is used because repeated experimental runs are expensive.
  • Statistical Testing: A repeated-trial sanity check reports that Binomial SD is a reasonable confidence-interval estimate when only one run is available.The comparison uses GPT-3.5 on iMEDQA and contrasts one-run Binomial SD with random-seed variation over five runs.

D Detailed Abstention Results

Abstention thresholds trade off additional questioning against diagnostic performance. Increasing the threshold generally improves performance by eliciting more questions, but excessive questioning leads to stagnation as questions become irrelevant or repetitive.

  • Threshold Analysis: Higher confidence thresholds generally make the Expert ask more questions and increase performance on iMEDQA.The threshold is selected through a grid search over the dataset.
  • Threshold Analysis: Performance stagnates when the number of follow-up questions becomes too high, potentially because later questions are irrelevant or repetitive.This identifies a practical trade-off between information gathering and interaction length.
  • Threshold Analysis: Figure 10 compares abstention strategies across increasing confidence thresholds, with separate encodings for rationale generation and self-consistency.The caption identifies Scale+RG+SC as the BEST system.

E Results on Additional LLMs

Additional experiments examine prompt robustness, model capacity, information availability, and question characteristics. Results indicate that interaction benefits some difficult specialties and questions, while smaller models and some specialties show less reliable or occasionally negative effects.

  • Additional LLMs: Llama-2-Chat models generally show behavior similar to stronger models, but limited capacity makes some settings unreliable and near chance.Interactive BASIC for Llama-2-7B is given as an example.
  • Information Availability: The information-availability comparison distinguishes a basic interactive prompt from the non-interactive initial setup and marks results that surpass it.This comparison is reported for Llama-2-Chat models.
  • Prompt Variations: Prompt variations produce broadly consistent Llama-2-Chat performance, with modest gains from rationale generation and no effect from answer-pair shuffling.The comparison covers system prompts and response formats.
  • Question Characteristics: Interactive information seeking is evaluated across medical specialties, demographics, and difficulty levels, with the largest benefits reported for ophthalmology and harder questions.Figure 11 organizes the analysis by specialties benefiting most or least and by difficulty.
  • Question Characteristics: 18.2% to 45.5%: Ophthalmology diagnostic accuracy improves after interaction.The paper presents this as an example of potential benefit in a specialty with initially low accuracy.
  • Question Characteristics: 43.4% to 48.3%: Accuracy improves for more challenging, clinically focused Step 2 and Step 3 questions.Family medicine and psychiatry sometimes show confusion after interaction, whereas neurosurgery and genomics show similar positive trends.
Loading 2406.00922v3…