Are LLMs Ready to Assist Physicians? PhysAssistBench for Interactive Doctor-Patient-EHR Assistance

2026-05-24 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Health & Medical Research · Depth: Expert, extended

Summary

PhysAssistBench is a new benchmark designed to evaluate large language models (LLMs) in interactive physician assistance scenarios, moving beyond isolated capability tests. Built from real MIMIC-IV cases, it uses a scalable multi-agent pipeline to create "agentic patients" that transform static EHR records into multi-turn clinical interactions while preserving clinical factuality. The benchmark features 324 multi-turn sessions and 1,296 manually reviewed and physician-validated turns, spanning 4 clinical scenarios (Diagnostic Workup, Med Safety, Treatment Response, Discharge Planning), 4 tasks (Information Lookup, Data Gathering, Clinical Reasoning, Write/Update), and 3 physician-query implicitness subtypes (Nominal Anaphora, Predicate Ellipsis, Abstract Event Anaphora). Experiments with 14 leading LLMs, including GPT-5.4, Claude-Opus-4.7, and GLM-5, reveal that current models remain unreliable. For instance, Claude-Opus-4.7 achieved 23.5% (EN) and 26.9% (ZH) Pass@Session at a 0.60 rubric score threshold, highlighting a key bottleneck in coordinating knowledge, communication, and system interaction.

Key takeaway

For AI Scientists and Machine Learning Engineers developing clinical LLMs, you must shift evaluation focus from isolated capabilities to integrated, multi-turn physician assistance. Your models currently struggle with coordinating knowledge, patient communication, and EHR tool use, especially with implicit queries and multi-tool composition. Prioritize developing robust interaction capabilities and addressing language-conditioned biases that inhibit tool invocation, as current models are unreliable for consistent, session-level clinical support.

Key insights

LLMs struggle with interactive physician assistance requiring coordinated knowledge, communication, and EHR system interaction.

Principles

• Physician assistance needs integrated capabilities, not isolated gains.
• Multi-turn interaction and implicit queries significantly degrade LLM performance.
• Sustaining quality across turns is a distinct, harder capability than turn-level accuracy.

Method

PhysAssistBench uses a multi-agent pipeline to transform MIMIC-IV records into interactive, record-grounded "agentic patients" for multi-turn doctor-patient-EHR scenarios, ensuring clinical factuality and scalability.

In practice

• Evaluate LLMs on integrated, multi-turn clinical workflows.
• Focus on improving multi-tool composition and handling implicit requests.
• Address language-conditioned safety priors that suppress EHR tool calls.

Topics

• Medical LLMs
• Physician Assistance
• EHR Interaction
• PhysAssistBench
• Multi-agent Systems
• Clinical Benchmarking
• FHIR R4

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Are LLMs Ready to Assist Physicians? PhysAssistBench for Interactive Doctor-Patient-EHR Assistance — Medical LLM assistance requires coordinated capabilities across knowledge, communication, and systems, not isolated gains.
• Why Specialist Models Still Matter: A Heterogeneous Multi-Agent Paradigm for Medical Artificial Intelligence — The future of medical AI lies in orchestrating generalist LLMs, specialist models, and clinicians through multi-agent collaboration.
• How good are ‘AI doctors’ — and will they take over medicine? — Advanced AI models are matching or exceeding human diagnostic accuracy in specific medical contexts, but real-world application is complex.
• Have LLMs improved patient outcomes? — Current LLMs show minimal evidence of directly improving patient or doctor health outcomes.
• Can AI Chatbots Reason Like Doctors? — Large language models demonstrate potential in clinical reasoning, but require careful integration and standardized evaluation in medical workflows.
• Enhancing Clinical Trial Patient Matching through Knowledge Augmentation and Reasoning with Multi-Agent — Multi-agent systems can augment LLMs with domain knowledge to improve clinical trial patient matching accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.