PaperlayerView exact revision on arXiv

Source-linked AI summary

HealthQ: Unveiling Questioning Capabilities of LLM Chains in Healthcare Conversations

Ziyu Wang, Hao Li, Di Huang, Hye-Sung Kim, Chae-Won Shin, Amir M. Rahmani

arXiv:2409.19487v4cs.CLcs.LG

TL;DR

Healthcare chains need to ask questions that elicit critical patient information, but this capability is under-explored relative to answering questions. HealthQ evaluates question quality and information gathered using LLM judges and NLP metrics; advanced chains outperform simpler baselines on reported summarization measures. Its evaluation still relies on LLMs and lacks clinician assessment in real clinical settings.

  • Problem

    Healthcare QA benchmarks emphasize answers, leaving the interactive ability of LLM chains to ask questions that elicit critical patient information under-explored.

    Source: §I · p. 1, C. Interactive Questioning with LLM… · p. 2

  • Method

    HealthQ evaluates healthcare-chain questions with LLM judges and NLP metrics, using custom datasets and multiple chain architectures.

    Source: §I · pp. 1–2

  • Results

    Advanced chains such as RAG Reflection and RAG Reflection CoT consistently outperform hardcoded and ReAct baselines across NER-based and ROUGE metrics.

    Source: C. Results and Analysis · p. 7

  • Takeaways & Limitations

    HealthQ provides a framework for evaluating question quality alongside the informativeness of patient responses across multiple judges and datasets.

    Source: §VI · p. 10

    Evaluation relies on LLMs and may miss aspects of question quality that healthcare professionals could assess; clinician evaluation in real clinical settings remains future work.

    Source: B. Evaluation Limitations · p. 10

Abstract

from arXiv · show

Effective patient care in digital healthcare requires large language models (LLMs) that not only answer questions but also actively gather critical information through well-crafted inquiries. This paper introduces HealthQ, a novel framework for evaluating the questioning capabilities of LLM healthcare chains. By implementing advanced LLM chains, including Retrieval-Augmented Generation (RAG), Chain of Thought (CoT), and reflective chains, HealthQ assesses how effectively these chains elicit comprehensive and relevant patient information. To achieve this, we integrate an LLM judge to evaluate generated questions across metrics such as specificity, relevance, and usefulness, while aligning these evaluations with traditional Natural Language Processing (NLP) metrics like ROUGE and Named Entity Recognition (NER)-based set comparisons. We validate HealthQ using two custom datasets constructed from public medical datasets, ChatDoctor and MTS-Dialog, and demonstrate its robustness across multiple LLM judge models, including GPT-3.5, GPT-4, and Claude. Our contributions are threefold: we present the first systematic framework for assessing questioning capabilities in healthcare conversations, establish a model-agnostic evaluation methodology, and provide empirical evidence linking high-quality questions to improved patient information elicitation.

I. INTRODUCTION

Digital healthcare uses AI and LLMs for tasks including patient management, virtual assistance, and medical question answering. HealthQ addresses the underexplored need to evaluate whether healthcare chains ask questions that elicit critical patient information.

B. LLMs in Healthcare

  • Healthcare LLM research has largely emphasized question answering, supported by benchmarks such as MedQA, emrQA, and MedMCQA.LLMs are also used in virtual assistants, automated diagnostic tools, and clinical reasoning systems.Source: §I · p. 1, B. LLMs in Healthcare · p. 2

I. INTRODUCTION

  • Patient care also requires chains to ask context-aware questions that elicit symptoms, medical history, and lifestyle factors, a capability that remains under-explored.Existing static, single-turn QA benchmarks do not assess these interactive information-seeking abilities.Source: §I · p. 1
  • HealthQ evaluates question quality and information elicitation using NLP measures such as ROUGE and NER alongside LLM-judged dimensions such as relevance and specificity.The framework uses custom datasets derived from ChatDoctor and MTS-Dialog and evaluates chains and judges across multiple LLMs.Source: §I · pp. 1–2

A. Digital Healthcare Interventions

  • Digital healthcare interventions use wearable devices and AI-driven analytics for remote monitoring, personalized treatment, and predictive analytics.Machine learning and natural language processing are also used for disease prediction, clinical-data interpretation, and decision support.Source: A. Digital Healthcare Interventions · p. 2

C. Interactive Questioning with LLM Healthcare Chains

Prior medical benchmarks prioritize answer accuracy, while sophisticated question-formulation strategies remain less studied. HealthQ evaluates healthcare-chain questions and their relation to the completeness of the information gathered.

C. Interactive Questioning with LLM Healthcare Chains

  • Medical QA benchmarks have advanced answer accuracy, but question formulation remains comparatively unexplored in interactive healthcare systems.CoT, reflective chains, and RAG can support contextual question generation, yet prior work has focused more on task completion and response accuracy.Source: C. Interactive Questioning with LLM… · p. 2

A. Overview of HealthQ

  • HealthQ benchmarks chains that simulate doctors using training data for retrieval and test data as virtual patients, without prior knowledge of test patients.The workflow generates patient statements, scores questions, and compares results with traditional NLP metrics.Source: A. Overview of HealthQ · p. 2
  • An LLM judge scores question quality, while summarization-based evaluation measures whether generated answers capture the patient case.The framework assesses questions for relevance, specificity, and informativeness and compares answers with detailed test-patient information.Source: A. Overview of HealthQ · p. 3
  • HealthQ analyzes relationships between question quality and answer comprehensiveness using correlation and mutual-information studies.Its workflow also evaluates question quality and answers with summary-based metrics such as NER2NER and ROUGE.Source: A. Overview of HealthQ · p. 3

B. Data

HealthQ reformats consultations into paired patient-known information and doctor information-seeking behavior, providing ground truth for evaluating generated questions. It uses separate training and evaluation data and tests on ChatDoctor and MTS-Dialog.

B. Data

  • Each consultation is represented by patient-known information and the information a doctor sought, enabling assessment of whether generated questions could uncover similar critical details.The processed format extracts patient knowledge and reformulates doctor questions from the original consultation.Source: B. Data · p. 3, Table I · p. 3

1) Dataset:

  • The evaluation uses ChatDoctor’s 110,000 anonymized medical conversations and MTS-Dialog’s 1,700 doctor-patient dialogues with clinical notes.Together, the datasets cover varied symptoms, diagnoses, treatment plans, and medical specialties.Source: 1) Dataset · p. 3

2) Data Processing:

  • Medical notes are divided into a knowledge-base set and a separate evaluation set, so chains query training references while encountering new patient scenarios during evaluation.The knowledge-base set is indexed in a vector database; this creates a searchable repository rather than traditional model training.Source: 2) Data Processing · p. 3
  • Evaluation notes are parsed into structured patient-known information and doctor-question statements, while knowledge-base entries pair content with metadata.The collate function creates these records, and the knowledge base supports medical-information retrieval during question generation.Source: 2) Data Processing · p. 4

C. Baseline Healthcare LLM Chains

The study compares healthcare chains that range from a fixed retrieval-and-template workflow to retrieval-augmented generation, reflection, iterative reasoning, and tool use. Their mechanisms differ in how questions are generated and refined.

C. Baseline Healthcare LLM Chains

  • The hardcoded workflow retrieves similar cases, extracts symptoms and medications with NER, then confirms them through predefined patient-facing templates.It uses embeddings and NER but does not require LLM inference.Source: C. Baseline Healthcare LLM Chains · p. 4
  • RAG retrieves relevant vector-database cases and uses a language model to generate concise, specific questions informed by the retrieved information.The study compares LangChain RetrievalQA with a customized version designed to generate questions rather than answers.Source: C. Baseline Healthcare LLM Chains · p. 4
  • RAG with Reflection [35] re-evaluates an initial question against retrieved cases and refines it for greater targeting and contextual fit.Source: C. Baseline Healthcare LLM Chains · p. 4
  • RAG with CoT iteratively reasons over retrieved context to refine an initial question or intermediate reasoning toward greater specificity and relevance.Source: C. Baseline Healthcare LLM Chains · p. 4
  • RAG with Reflection and CoT-Self-Consistency [39] compares results from multiple reasoning branches with an LLM to produce a refined, consistent question.Unlike the original voting approach, this implementation uses an LLM to compare branch results.Source: C. Baseline Healthcare LLM Chains · pp. 4–5
  • ReAct [40] lets an LLM decide whether to call a tool, then incorporates the executor’s output into its next input.The implementation offers vector-database search and medical-abbreviation lookup as separate tools.Source: C. Baseline Healthcare LLM Chains · p. 5

D. Evaluation Framework

The framework evaluates both generated-question quality and how much patient information the questions elicit, using a simulated patient, an LLM judge, and summarization metrics.

  • The virtual patient uses patient knowledge K to generate an initial statement and answers to questions, with topic and factual content controlled by K.The authors identify LLM phrasing and inconsistency with K as sources of bias.Source: D. Evaluation Framework · p. 5
  • An LLM judge scores each generated question against patient knowledge using specificity, usefulness, relevance, coverage, and fluency.Source: D. Evaluation Framework · p. 5
  • ROUGE and NER-based evaluation assess how well simulated answers capture the patient information represented in the test data.The process compares generated answers with comprehensive patient information.Source: D. Evaluation Framework · p. 5
  • Healthcare chains receive only the patient statement, while judges can access additional ground truth; chains do not receive test ground truth.Tools may be derived from training sets or external sources, but not the test ground truth.Source: D. Evaluation Framework · p. 6

E. Validation

Validation examines dependencies between question ratings and the completeness of elicited answers, averaging metrics across interactions for each chain and using normalized mutual information (NMI).

  • NMI measures dependencies between aggregated interrogation and summarization metrics across chains, with scores averaged over interactions for each chain.The aggregate analysis is intended to capture systematic patterns rather than isolated cases or outliers.Source: E. Validation · p. 6
  • NMI normalizes mutual information to make dependencies comparable across metrics, while scores are linearly rescaled for visualization.The framework separates question generation from evaluation and prevents chains from accessing test ground truth.Source: E. Validation · p. 6
  • NMI ranges from 0 for no dependency to 1 for perfect association, and captures linear and nonlinear relationships between metrics.A high NMI between Relevance and ROUGE-L Recall indicates that highly relevant questions are more likely to elicit comprehensive answers aligned with ground truth.Source: E. Validation · p. 6
  • The validation asks whether induced answers summarize relevant information and whether higher-rated questions are associated with more comprehensive answers.It uses summarization metrics and NMI to assess these relationships.Source: E. Validation · p. 6

A. Experiment Setup

The experiments benchmark healthcare chains with multiple LLM judges and combine question ratings, medical-entity comparisons, and ROUGE measures against patient-note ground truth.

A. Experiment Setup

  • Mixtral served on Groq is the primary chain model, with GPT-3.5-turbo, GPT-4, and Claude-3-Opus used as judges.Open-source models are used where possible; CoT-SC uses temperature 0.3, while other models use 0.Source: A. Experiment Setup · pp. 6–7
  • The setup uses FAISS for retrieval and MiniLM embeddings, with Claude-3-Opus, GPT-3.5-turbo, and GPT-4 used for preprocessing tasks.Source: A. Experiment Setup · p. 7
  • The study evaluates 128 test cases, benchmarks seven LLM healthcare chains, and generates 796 patient status–doctor statement pairs.Cases come equally from the two datasets, and results are grouped by chain type and judge model.Source: A. Experiment Setup · p. 7

B. Evaluation Metrics Implementation

  • Question quality is scored from 0 to 1 on Relevance, Usefulness, Specificity, Coverage, and Fluency by GPT-3.5, GPT-4, and Claude-3-Opus.Higher scores indicate better performance.Source: B. Evaluation Metrics Implementation · p. 7
  • Biomedical NER compares medical entities in generated answers with ground-truth patient data, while ROUGE measures overlap between elicited information and comprehensive ground truth.ROUGE-1 measures word overlap, ROUGE-2 two-word phrase overlap, and ROUGE-L longest matching subsequences.Source: B. Evaluation Metrics Implementation · p. 7
  • Original test-set patient notes provide the gold standard, including symptoms, medical history, and contextual details.The combined judge, NER, and ROUGE metrics assess question quality and the informativeness of elicited answers.Source: B. Evaluation Metrics Implementation · p. 7

C. Results and Analysis

Advanced reflective chains outperformed simpler baselines on question and answer metrics. Aggregated metric associations were strongest for relevance and usefulness, while fluency and coverage showed weaker links to answer quality.

C. Results and Analysis

  • RAG Reflection and RAG Reflection CoT outperformed simpler chains on interrogation metrics; GPT-4o specificity was 7.98 for RAG Reflection, while fluency was 9.33 for RAG Reflection CoT.RAG Reflection's GPT-4o coverage score was 5.39, compared with 3.69 for the hardcoded chain.Source: Figure 2 · p. 8, C. Results and Analysis · p. 8
  • RAG Reflection's ROUGE-L scores were 0.70 for GPT-4o, 0.72 for GPT-3.5, and 0.67 for Claude, alongside high relevance and specificity across judges.The radar plots show larger, more balanced areas for RAG Reflection and RAG Reflection CoT than for ReAct and hardcoded chains.Source: C. Results and Analysis · p. 8
  • RAG Reflection outperformed hardcoded and ReAct chains across NER-based and ROUGE metrics, scoring NER2NER total 0.65 and ROUGE-L F 0.62 under GPT-4o.RAG Reflection also led on these measures under GPT-3.5 and Claude.Source: C. Results and Analysis · p. 7, Table IV · p. 9
  • Fluency–NER2NER med (NMI = 0.66) and Coverage–ROUGE-2 Precision (NMI = 0.53) were weaker dependencies.The reported patterns indicate that linguistic clarity need not correspond to detailed entity extraction, and broad coverage need not yield precise responses.Source: C. Results and Analysis · p. 9
  • Specificity aligned with ROUGE-1 Recall (NMI = 0.79) and NER2NER total (NMI = 0.71).The results describe specificity as important for targeted, informative interactions.Source: C. Results and Analysis · p. 9
  • Aggregated NMI linked Relevance with ROUGE-L Recall (0.85) and Usefulness with NER2NER total (0.84), consistent with broader Relevance and Usefulness alignment to ROUGE-L.These associations connect LLM-judge question ratings with traditional summarization measures.Source: Figure 3 · p. 9, C. Results and Analysis · p. 9

A. Sources of Bias and Correction

HealthQ’s bias discussion covers sampling and data-construction risks, LLM judgment biases, and the limitations of LLM-based evaluation. The conclusion reports consistent performance across judges and datasets and identifies further evaluation directions.

A. Sources of Bias and Correction

1) Pre-judgement Errors:

  • Randomized splits and sampling may create distributional bias; cleaning hallucinations or incomplete case data may omit diagnostic information, despite assumptions of comprehensive, realistic patient data.The paper suggests patient profiling and matching or fixed conversation starters as possible mitigations.Source: 1) Pre-judgement Errors · pp. 9–10

2) Judgement Errors:

  • Prompting and judgment errors can bias interrogation, while LLM-simulated answers can bias summarization metrics; the proposed mitigation diversifies LLM roles.Source: 2) Judgement Errors · p. 10

B. Evaluation Limitations

  • LLM-based evaluation may miss aspects of question quality that only healthcare professionals can assess; future work proposes clinician review, real clinical testing, and longitudinal studies.The proposed studies would examine utility across specialties and alignment with diagnostic protocols.Source: B. Evaluation Limitations · p. 10

VI. CONCLUSION

  • HealthQ showed consistent performance across judge models and datasets; future work will expand to multi-turn interactions, refine metrics, and explore broader healthcare applications.The conclusion also describes studies linking question quality with answer informativeness.Source: §VI · p. 10

Related papers

Loading 2409.19487v4…