Electronic Health Records (EHRs) contain a vast wealth of patient data, but extracting actionable insights from them is notoriously difficult due to their unstructured, jargon-heavy nature. Early paradigms in clinical question answering (QA) focused on mapping natural language queries to structured databases (Preethi Raghavan et al. in emrKBQA: A Clinical Knowledge-Base Question Answering Dataset) or using multi-task learning to classify question intent and recognize medical entities (Turdi Tohti et al. in Medical QA Oriented Multi-Task Learning Model for Question Intent Classification and Named Entity Recognition). To handle the sheer variety of ways a user might phrase a query, researchers also developed paraphrase generation systems to expand question coverage (Sarvesh Soni et al. in A Paraphrase Generation System for EHR Question Answering).
However, as patient portals become more widespread, the demand has shifted from structured database lookups to answering complex, free-text patient questions directly from unstructured clinical notes (Sarvesh Soni et al. in Overview of the ArchEHR-QA 2025 Shared Task on Grounded Question Answering from Electronic Health Records). This shift introduces a critical requirement: answers must be grounded in explicit evidence from the patient's record to ensure safety and verifiability.
The Core Challenge: Hallucination vs. Verifiability
Why can we not simply pass an entire EHR into a modern Large Language Model (LLM) and ask it to generate an answer? The primary obstacle is factuality. In medicine, an ungrounded or hallucinated response can lead to severe clinical mismanagement. To build trust, a system must provide a concise, professional answer while precisely citing the exact sentences within the clinical notes that support the claim (Proceedings of the 24th Workshop on Biomedical Language Processing (Shared Tasks)).
Furthermore, raw EHR data is noisy and redundant. Throwing more text at a model does not necessarily yield better summaries; in fact, Erlend Frayling et al. demonstrated that isolating a single, highly relevant section (such as the "History of Present Illness") often produces superior clinical summaries compared to combining multiple disparate sections (UoG Siephers at “Discharge Me!”: Exploring Ways to Generate Synthetic Patient Notes From Multi-Part Electronic Health Records). Similarly, consumer queries are frequently rambling or poorly formulated, requiring preprocessing or focus-extraction techniques to isolate the core question before attempting retrieval (Jooyeon Lee et al. in MNLP at MEDIQA 2021: Fine-Tuning PEGASUS for Consumer Health Question Summarization).
The Two-Stage Architecture: Locate then Synthesize
To guarantee that an answer is strictly evidence-based, modern architectures decouple the problem into a two-stage pipeline: Evidence Retrieval (locating the exact supporting sentences) and Grounded Generation (synthesizing the final answer with explicit citations). This is conceptually similar to multi-evidence natural language inference (NLI) frameworks used to verify claims against clinical trial reports (Mingtong Huang et al. in CPIC at SemEval-2023 Task 7: GPT2-Based Model for Multi-evidence Natural Language Inference for Clinical Trial Data).
[Patient Question] + [EHR Notes]
│
▼
┌──────────────────────────────┐
│ Stage 1: Evidence Retrieval │ ──► Identifies exact supporting sentences
└──────────────────────────────┘
│
▼
┌──────────────────────────────┐
│ Stage 2: Grounded Generation │ ──► Synthesizes concise answer with citations
└──────────────────────────────┘
│
▼
[Verified, Traceable Answer]
Stage 1: Evidence Selection Strategies
Finding the exact "needle" of evidence in a lengthy clinical note requires highly precise retrieval mechanisms. Researchers have explored several distinct strategies to optimize this stage:
- Fine-Tuned Embeddings: Arina Zemchyk utilizes contrastive fine-tuning on embedding models to maximize the semantic alignment between patient questions and relevant clinical sentences (razreshili at ArchEHR-QA 2025: Contrastive Fine-Tuning for Retrieval-Augmented Biomedical QA).
- Multi-Stage Re-ranking: Provia Kadusabe et al. implement a two-stage pipeline that first retrieves a broad set of candidate sentences and then re-ranks them to ensure only the most factual context is passed to the generator (DKITNLP at ArchEHR-QA 2025: A Retrieval Augmented LLM Pipeline for Evidence-Based Patient Question Answering).
- Prompt Chaining: In contrast to training specialized embeddings, Adam Remaki et al. show that prompting LLMs via structured "prompt chains" can actually outperform fine-tuned embeddings at identifying essential medical information (LIMICS at ArchEHR-QA 2025: Prompting LLMs Beats Fine-Tuned Embeddings).
- Test-Time Scaling: To bypass the need for large, task-specific annotated datasets, Tuan Dung Le et al. introduce a test-time scaling approach, leveraging the inherent reasoning capabilities of general-purpose LLMs to dynamically select evidence (LAILab at ArchEHR-QA 2025: Test-time scaling for evidence selection in grounded question answering from electronic health records).
Stage 2: Faithful Synthesis and Citation Compliance
Once the evidence sentences are isolated, the generative model faces a dual constraint: it must translate complex medical jargon into a concise, patient-friendly response while strictly adhering to formatting rules and in-text citations.
This multi-objective task is notoriously difficult to control. For instance, the author of the razreshili system noted that while generative models excel at relevance, they frequently struggle with strict word count limits and precise citation formatting (razreshili at ArchEHR-QA 2025: Contrastive Fine-Tuning for Retrieval-Augmented Biomedical QA). To overcome this, Sai Prasanna Teja Reddy Bogireddy et al. employ agentic prompt optimization (such as MIPROv2) to automatically refine instructions, significantly boosting both evidence extraction accuracy and citation faithfulness (Neural at ArchEHR-QA 2025: Agentic Prompt Optimization for Evidence-Grounded Clinical Question Answering).
Crucially, this pipeline does not require massive computing infrastructure. In resource-constrained environments, the authors of recent systems have shown that competitive performance can be achieved by leveraging smaller, open-source language models through relevance classification ensembles (Sebastian Balmus et al. in UniBuc-SB at ArchEHR-QA 2025: A Resource-Constrained Pipeline for Relevance Classification and Grounded Answer Synthesis) or multi-prompt ensembling techniques (Dragos Ghinea et al. in UNIBUC-SD at ArchEHR-QA 2025: Prompting Our Way to Clinical QA with Multi-Model Ensembling).
Why Grounded QA Works
The intuition behind the success of the two-stage grounded QA model lies in how it constrains the language model's role. Instead of treating the LLM as a knowledge retrieval engine—which triggers its tendency to hallucinate plausible-sounding medical facts—the architecture treats the LLM as a constrained reasoning engine.
By forcing the system to first commit to a discrete set of source sentences, the generation phase is bounded. The model's task is reduced to synthesizing and translating the pre-selected clinical evidence into an accessible format. If a claim cannot be mapped back to an explicit citation index, it is excluded, providing a verifiable audit trail that ensures clinical safety and patient trust.
Key Foundational Work
A Paraphrase Generation System for EHR Question Answering (2019) — contributed 1 concept elements to this direction.
emrKBQA: A Clinical Knowledge-Base Question Answering Dataset (2021) — contributed 1 concept elements to this direction.
MNLP at MEDIQA 2021: Fine-Tuning PEGASUS for Consumer Health Question Summarization (2021) — contributed 1 concept elements to this direction.