From Generation to Collaboration: Using LLMs to Edit for Empathy in Healthcare

TL;DR

The paper tackles the challenge of providing empathetic yet factually accurate clinical communication by repurposing LLMs as empathy editors rather than autonomous generators. It introduces a novel editing framework with simple and refined prompts, plus two evaluation tools—3EM-Ranker for empathy and MedFactChecking for factuality—validated against human judgments and clinician feedback. Across a 163-pair prostate cancer QA dataset, editors consistently increase perceived empathy while preserving medical facts, with Gemini-2.5-Flash delivering the strongest factual grounding. The findings support deploying LLMs as collaborative editorial assistants to enhance compassionate communication in healthcare while mitigating hallucinations and unsafe content, and they outline practical trade-offs between empathy intensity and factual integrity.

Abstract

Clinical empathy is essential for patient care, but physicians need continually balance emotional warmth with factual precision under the cognitive and emotional constraints of clinical practice. This study investigates how large language models (LLMs) can function as empathy editors, refining physicians' written responses to enhance empathetic tone while preserving underlying medical information. More importantly, we introduce novel quantitative metrics, an Empathy Ranking Score and a MedFactChecking Score to systematically assess both emotional and factual quality of the responses. Experimental results show that LLM edited responses significantly increase perceived empathy while preserving factual accuracy compared with fully LLM generated outputs. These findings suggest that using LLMs as editorial assistants, rather than autonomous generators, offers a safer, more effective pathway to empathetic and trustworthy AI-assisted healthcare communication.

Figures (12)

• Figure 1: AI editing versus generation for clinical responses. Direct AI generation produces empathetic but factually inaccurate responses with hallucinated medical details, while AI editing of physician responses maintains factual accuracy while adding empathy.
• Figure 1: Editing prompt for empathy enhancement. PQ refers to the patient question and PR refers to the physician's response.
• Figure 2: Bidirectional fact-checking framework for measuring factual accuracy in AI-edited clinical responses. Fact-Recall (green arrow) quantifies information loss from the original physician response, while Fact-Precision (yellow arrow) detects hallucinated additions in the edited response. Red text indicates facts that fail entailment checking: lost information (left) and unsupported additions (right).
• Figure 2: The refined editing prompt with explicit behavioral constraints.
• Figure 3: MedFactChecking Score analysis across models. Top: Micro (left) and macro (right) metrics for recall, precision, and F1. Bottom: Fact flow showing preserved and grounded facts (left); fact count distributions (right). Gemini-2.5-Flash substantially outperforms other models.