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 enhancing communication, current approaches face significant limitations:

1. Autonomous Generation Risks: End-to-end AI generation often produces empathetic responses that suffer from factual hallucinations, introducing medical inaccuracies or "phony empathy."
2. Fine-tuning Trade-offs: Fine-tuning LLMs for empathy can inadvertently degrade factual reliability, leading to sycophantic validation or the dilution of critical clinical guidance.
3. Measurement Gaps: Existing empathy metrics (e.g., Jefferson Scale) are designed for human interactions, not algorithmic text, and automated empathy evaluation often lacks alignment with human judgment.

The core challenge is to enhance the empathetic tone of physician responses without compromising the factual integrity of the medical information provided.

Methodology

1. Proposed Approach: Collaborative Editing

Instead of using LLMs as autonomous generators, the authors propose a collaborative editing framework. In this workflow:

• The physician provides an initial, factually accurate but potentially terse response.
• An LLM acts as an editor, refining the text to enhance empathy while strictly preserving the original medical facts.
• Two prompt strategies are tested:
  • Simple Editing Prompt: Directs the model to revise for empathy while maintaining original meaning.
  • Refined Editing Prompt: Adds explicit behavioral constraints to prevent speculative statements and ensure clinical grounding.

2. Evaluation Framework

The study introduces two novel quantitative metrics to systematically assess the trade-off between empathy and factuality:

• 3EM-Ranker (Empathy Evaluation):

  • An improvement over the binary EM-Rank framework.
  • Utilizes a three-way comparison (Response A > Response B, Response B > Response A, or Equally Empathetic).
  • Validated against human patient annotators, showing higher alignment (Pearson correlation ~0.55–0.57) than previous binary models, addressing the subjectivity of empathy.

• MedFactChecking Score (Factuality Evaluation):

  • A bidirectional entailment framework adapted from FactEHR.
  • Fact-Recall: Measures the proportion of original medical facts preserved in the edited response (detecting information loss).
  • Fact-Precision: Measures the proportion of facts in the edited response that are grounded in the original physician response (detecting hallucinated additions).
  • Uses bidirectional entailment checks to separately quantify omission and hallucination errors.

3. Dataset and Models

• Dataset: 163 patient-physician message pairs from a urology patient portal (prostate cancer post-prostatectomy). Data was manually de-identified.
• Models Evaluated: Five LLMs (Qwen3-7/14B, Mistral-7B, Llama-3.1-8B, Gemini-2.5-Flash).
• Validation: Expert validation by board-certified urologists and ML researchers to verify fact extraction accuracy and clinical impact of errors.

Key Results

Empathy Enhancement (RQ1)

• LLM Superiority in Tone: LLM-generated responses (both direct and edited) were consistently judged as more empathetic than original physician responses (>90% of cases).
• Prompt Constraints: Stricter prompts (Refined vs. Simple) resulted in slightly lower empathy rankings but significantly improved factual safety.
• Human Alignment: The 3EM-Ranker showed strong alignment with human patient judges, particularly in identifying "equal" empathy cases where binary models failed.

Factual Preservation (RQ2)

• Model Performance: Gemini-2.5-Flash significantly outperformed other models, achieving a micro-recall of 0.92 and micro-precision of 0.91. Other models (LLaMA, Mistral, Qwen) showed lower precision (0.77–0.79 F1), introducing more hallucinated content.
• Length Dependency: Short physician responses (<200 characters) were most vulnerable to hallucination in non-Gemini models, which attempted to "fill in" missing context with ungrounded facts. Gemini-2.5-Flash maintained stable precision across all response lengths.
• Error Patterns: Expert review identified specific failure modes in edited responses, including unprompted follow-up recommendations, clinical assumptions, and false assurance.

The Empathy-Factuality Trade-off (RQ3)

• Inverse Relationship: Increasing empathy intensity (from "Empathic" to "Extreme Empathic") led to a 13-percentage-point drop in precision (from 0.91 to 0.78), primarily driven by hallucinated additions rather than information loss.
• Editing vs. Generation:
  • Refined Editing: Achieved near-perfect factual preservation (Recall ~0.99, Precision ~0.93) while improving empathy.
  • Direct Generation: Failed catastrophically on factual grounding (Recall ~0.39), preserving fewer than 40% of the facts a physician would include.
• Ranking Tension: A fundamental inverse relationship exists: Direct Generation > Simple Editing > Refined Editing > Physician Response for empathy, but Physician Response > Refined Editing > Simple Editing > Direct Generation for factuality.

Significance and Claims

The paper claims that reframing LLMs from autonomous generators to collaborative editors offers a safer, more effective pathway for AI in healthcare.

• Safety First: The study demonstrates that editing physician responses preserves the "source of truth" (the physician's medical facts) while adding emotional resonance, avoiding the hallucinations inherent in end-to-end generation.
• Practical Viability: The "Refined Editing" approach achieves the most favorable trade-off, offering a deployment-ready strategy that enhances patient communication without compromising clinical accuracy.
• Measurement Innovation: The introduction of the 3EM-Ranker and MedFactChecking Score provides the first systematic, automated framework to quantify both the emotional and factual quality of clinical AI interactions, addressing a critical gap in current NLP research.

The authors conclude that while empathy and factuality are often in tension, an editing-based approach allows clinicians to leverage AI for emotional support without surrendering medical precision, representing a crucial step toward responsible AI integration in healthcare.