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

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

Problem Statement
Clinical empathy is vital for patient trust and outcomes, yet physicians often struggle to balance emotional warmth with factual precision due to cognitive load and time constraints. While Large Language Models (LLMs) offer potential for healthcare communication, current approaches face a critical dichotomy:

1. Autonomous Generation: End-to-end LLM generation often produces highly empathetic responses but suffers from "hallucinations," introducing factual errors and speculative medical advice that compromise patient safety.
2. Fine-tuning: Training models specifically for empathy requires substantial resources and can inadvertently increase factual errors or create "phony empathy."
3. The Gap: There is a lack of systematic methods to enhance the empathy of physician-written responses without compromising the clinical accuracy of the original medical information.

Methodology
The authors propose a paradigm shift from autonomous generation to collaborative editing. Instead of generating responses from scratch, LLMs are used to edit existing physician responses to enhance empathetic tone while preserving the original medical facts.

• Data: A dataset of 163 patient-physician message pairs from a urology patient portal (prostate cancer patients) was used. All data underwent rigorous de-identification.
• Editing Strategy: Two prompt variations were tested:
  • Simple Editing: Instructs the model to revise for empathy while maintaining original meaning.
  • Refined Editing: Adds explicit behavioral constraints to prevent factual drift and speculative additions.
• Evaluation Framework: The study introduces two novel quantitative metrics:
  1. 3EM-Ranker (Empathy): A three-way comparison metric (Response A > Response B, Response B > Response A, or Equal) designed to better align with human judgment than binary ranking systems. It was validated against human patient annotators, achieving alignment scores of 0.55–0.57, significantly higher than prior work.
  2. MedFactChecking Score (Factuality): A bidirectional entailment framework adapted from FactEHR. It measures:
     • Fact-Recall: The proportion of original medical facts preserved in the edited response.
     • Fact-Precision: The proportion of facts in the edited response that are grounded in the original (detecting hallucinations).

Key Results
Experiments were conducted across five state-of-the-art LLMs (including Qwen3, Mistral, Llama-3, and Gemini-2.5-Flash).

• Empathy vs. Factuality Trade-off:
  • Direct Generation: Produces the highest perceived empathy but fails catastrophically on factuality (Micro Recall ~0.39), preserving fewer than 40% of the facts a physician would include.
  • Refined Editing: Achieves the optimal balance. Using Gemini-2.5-Flash, the refined editing approach preserved 99% of original facts (Micro Recall) and maintained 93% precision, while significantly improving empathy scores compared to raw physician responses.
  • Simple Editing: Showed modest degradation in factuality compared to refined editing, indicating that unconstrained empathy instructions lead to information loss and hallucinations.
• Model Performance: Gemini-2.5-Flash significantly outperformed open-source models in factual preservation during editing, particularly for short physician responses where other models tended to "fill in" missing context with hallucinated details.
• Empathy Intensity: Increasing the intensity of empathy instructions ("High" or "Extreme") led to a sharp decline in factuality (precision dropped from 0.91 to 0.78), primarily due to the addition of unsupported content rather than the loss of original facts.

Key Contributions

1. Paradigm Shift: The paper reframes the role of LLMs in healthcare from autonomous generators to "editorial assistants," prioritizing augmentation of human expertise over replacement.
2. Novel Metrics: Introduction of the 3EM-Ranker for nuanced empathy evaluation and the MedFactChecking Score for bidirectional fact verification (Recall and Precision) in edited clinical text.
3. Empirical Evidence: Demonstration that constrained editing preserves factual integrity (93–99% preservation) while meaningfully enhancing empathy, whereas direct generation sacrifices safety for tone.
4. Validation: Rigorous validation of metrics against human annotators (patients and urologists), showing that the proposed framework captures clinically significant errors and aligns better with human preference than existing automated methods.

Significance
The authors argue that using LLMs as collaborative editors offers a safer, more effective pathway for AI-assisted healthcare communication than autonomous generation. By maintaining the physician's factual intent while refining the emotional delivery, this approach addresses the tension between clinical accuracy and compassionate care. The findings suggest that for clinical deployment, systems must prioritize factual grounding over maximum empathetic expression, utilizing constrained editing workflows to ensure trustworthiness and patient safety. The work represents a step toward responsible AI integration where technology supports, rather than supplants, clinical judgment.