Balancing Efficiency and Empathy: Healthcare Providers' Perspectives on AI-Supported Workflows for Serious Illness Conversations in the Emergency Department

This paper presents findings from interviews with 11 emergency department providers, revealing that while AI-supported workflows could address critical barriers to Serious Illness Conversations by synthesizing information and automating documentation, successful implementation requires design guidelines that balance efficiency gains with the preservation of human connection and clinical autonomy.

The Gist

Imagine you are a doctor in a busy Emergency Room (ER). It's chaotic. People are rushing in, machines are beeping, and you have to make life-or-death decisions in minutes. Now, imagine a patient walks in who is very sick, perhaps with heart failure or advanced cancer. You need to have a very serious, delicate conversation with them or their family: "What matters most to you? Do you want to fight for every possible moment, even if it means tubes and machines? Or do you want to focus on being comfortable and pain-free?"

These are called Serious Illness Conversations (SICs). They are crucial, but in the ER, they almost never happen. Why? Because the ER is like a hurricane, and these conversations require a calm, quiet room.

This paper is a study by researchers who interviewed 11 ER doctors and nurses to understand why these conversations are so hard to have, and how Artificial Intelligence (AI) could act as a helpful sidekick to make them possible without ruining the human connection.

Here is the breakdown of their findings, explained with simple analogies:

1. The Problem: The "Four-Stage" Bottleneck

The researchers found that having a serious conversation in the ER isn't just one step; it's a four-step process that breaks down at every turn. Think of it like trying to bake a complex cake while standing in a wind tunnel.

• Stage 1: Finding the Right People (Identification)

  • The Struggle: Doctors need to know if a patient has already discussed their wishes. But the patient's medical records are scattered like leaves in a storm. One note is in a folder here, another in a different system there. Doctors spend precious minutes hunting for a "Do Not Resuscitate" order or a note about the patient's values, often finding nothing or conflicting info.
  • The Result: They miss the chance to have the conversation because they can't find the map.

• Stage 2: Getting Ready (Preparation)

  • The Struggle: Even if they find the patient, the doctor has to mentally prepare for a heavy, emotional talk. But in the ER, they are often tethered to the bedside, managing other emergencies. They might be tired, stressed, or even dealing with their own grief (like a doctor whose father just died). They have no time to "warm up" for the conversation.
  • The Result: They walk into the room feeling unprepared and panicked.

• Stage 3: Having the Talk (Conduction)

  • The Struggle: This is the hardest part. The patient might be in shock, the family might be crying, and the doctor has to break bad news and ask deep questions in 5 minutes. It's like trying to negotiate a peace treaty while a building is on fire. There is no time to build trust, and the doctor often feels awkward or unsure of what to say next.
  • The Result: The conversation feels rushed, mechanical, or gets avoided entirely.

• Stage 4: Writing it Down (Documentation)

  • The Struggle: After the emotional talk, the doctor has to type up a long, detailed note. But they are exhausted and behind on other patients. They often stay late, unpaid, just to finish the paperwork. Plus, they write these notes in different places, so the next doctor (in the ICU or a nursing home) can't find them.
  • The Result: The hard work of the conversation is lost because the notes are messy or missing.

2. The Solution: AI as a "Silent Co-Pilot"

The doctors didn't want a robot to take over the conversation. They didn't want an AI to talk for them. Instead, they wanted AI to be a Silent Co-Pilot or a Super-Organized Assistant that works in the background.

Here is how they imagine AI helping, using the paper's design ideas:

• The "Flashlight" for Information:
  Instead of digging through 50 different files, the AI acts like a flashlight. It instantly gathers all the scattered notes, summarizes them, and says: "Here is what we know: The patient is a Red Sox fan, they have a daughter named Maria, and they previously said they hate being on machines. But we are missing a clear 'Do Not Resuscitate' order." This saves the doctor time so they can focus on the patient.

• The "Whispering Guide" in the Room:
  Imagine the doctor is talking to a crying family. The AI listens (with permission) and, on a small screen the doctor can glance at, it whispers: "The patient just mentioned 'comfort' three times. Maybe ask what 'comfort' means to them?" or "They seem scared of pain; maybe reassure them about pain management."

  • Crucial Point: The AI never interrupts. It just offers a gentle nudge, like a stage manager handing a cue card to an actor, so the doctor doesn't freeze up.

• The "Auto-Writer" for Notes:
  After the talk, the AI listens to the conversation and writes the first draft of the medical note. It highlights the important parts: "Patient chose comfort care because they value being pain-free over being alive on a machine." The doctor just reads it, tweaks it, and hits send. This saves hours of typing.

• The "Coach" for Learning:
  After the shift, the AI could give the doctor a gentle, private message: "You did a great job explaining the options. Next time, you might try asking about their grandchildren earlier to build a connection." This helps doctors get better at these hard talks without feeling judged.

3. The Big Balance: Efficiency vs. Empathy

The most important finding of the paper is a paradox.

• Efficiency (doing things fast) usually feels cold and robotic.
• Empathy (caring deeply) usually takes time and slows things down.

The doctors said: "We need efficiency to have empathy."

If the AI handles the boring stuff (finding files, typing notes, organizing data), the doctor is freed up to look the patient in the eye, hold their hand, and truly listen. The AI doesn't replace the human heart; it clears the clutter so the human heart can work better.

Summary

This paper argues that in the chaotic Emergency Room, we can't expect doctors to have perfect, deep conversations about life and death without help. By using AI as a background helper that organizes information, suggests gentle conversation starters, and handles the paperwork, we can give doctors the time and mental space they need to be truly human, compassionate, and present for their patients in their final moments.

The takeaway: AI shouldn't be the voice in the room; it should be the quiet hand that clears the table so the doctor can have a real conversation.

Technical Summary

1. Problem Statement

Serious Illness Conversations (SICs) are structured discussions regarding values, goals, and care preferences for patients with life-threatening illnesses. While evidence suggests that early SICs improve care alignment and reduce unnecessary interventions, they are rarely conducted in Emergency Departments (EDs).

The ED presents a unique, high-stakes environment where SICs are most needed but face significant barriers:

• Time Pressure: Decisions must be made within minutes, often during acute crises.
• Fragmented Information: Patient history and prior care preferences are scattered across Electronic Health Records (EHRs) or non-existent.
• Lack of Rapport: Providers often meet patients for the first time under crisis conditions, making it difficult to build the trust required for sensitive discussions.
• Emotional Burden: Providers face high cognitive and emotional loads without adequate training or support.
• Documentation Burden: Recording outcomes is time-consuming and often inconsistent, hindering continuity of care.

Current technological solutions are largely designed for low-pressure, longitudinal settings (e.g., oncology clinics) and fail to address the urgency, emotional intensity, and rapid decision-making demands of the ED. There is a critical gap in understanding how Artificial Intelligence (AI) can support ED providers in conducting SICs without compromising the essential human connection.

2. Methodology

The study employed a qualitative, user-centered research approach to investigate the current workflow and identify opportunities for AI integration.

• Participants: 11 healthcare providers (9 physicians, 2 nurses) working in tertiary academic medical centers in the United States. The sample included diverse roles (e.g., emergency medicine, palliative care, intensivist) and experience levels (4 to 42 years).
• Data Collection: Semi-structured interviews were conducted via Microsoft Teams. Participants were asked to describe recent SIC experiences, detailing their workflow, challenges, and perspectives on potential AI support.
• Data Analysis: Researchers used thematic analysis on interview transcripts.
  • Workflow Synthesis: Identified a four-stage workflow specific to EDs.
  • Coding: Two researchers performed open coding to generate initial codes, which were clustered into high-level thematic groups using affinity diagramming.
  • Rigor: The study followed trustworthiness criteria (iterative codebook development, negotiated agreement) and reported theme prevalence using qualitative terminology (e.g., "most," "some").

3. Key Contributions

The paper makes three primary contributions to the fields of HCI, healthcare informatics, and AI:

1. A Novel Four-Phase ED-Specific SIC Workflow: The study synthesizes a distinct workflow for conducting SICs in the ED, consisting of:
   • Identification: Determining if an SIC is needed.
   • Preparation: Mental and logistical planning.
   • Conduction: The actual conversation with the patient/family.
   • Documentation: Recording outcomes in the EHR.
2. Empirical Analysis of Barriers and Needs: A detailed mapping of specific challenges (e.g., fragmented data, lack of training, emotional distress) and provider needs at each stage of the workflow.
3. Design Guidelines for Ambient AI: A set of four design guidelines for integrating AI into ED workflows that balance efficiency (reducing cognitive load) with empathy (preserving human connection).

4. Key Results & Findings

A. The Four-Stage Workflow and Challenges

1. Identification:
   • Challenge: Providers struggle to find relevant data (code status, goals) in fragmented EHRs within minutes. Keyword searches often fail, and documentation is inconsistent.
   • Need: Rapid synthesis of actionable patient information.
2. Preparation:
   • Challenge: Providers face "panic" due to lack of time, insufficient training, and the emotional weight of sudden crises. They often lack a "patient story" to build rapport.
   • Need: Tools to synthesize a coherent patient narrative and provide emotional/logistical scaffolding.
3. Conduction:
   • Challenge: Initiating sensitive topics with unprepared patients/families in noisy, crowded spaces. Providers fear breaking the "human connection" if they rely too heavily on technology.
   • Need: Real-time, contextual guidance that is ambient and peripheral (not intrusive) to help navigate the conversation without displacing empathy.
4. Documentation:
   • Challenge: Time-consuming, unstructured note-taking often done after shifts. Current notes lack the "rationale" behind decisions, hindering continuity for other teams.
   • Need: Automated generation of rationale-driven notes that capture why a decision was made, not just what was decided.

B. The "Efficiency-Empathy Paradox"

A central finding is the tension providers feel: they need AI to handle the cognitive burden (efficiency) to free up mental space for emotional engagement (empathy), yet they fear AI will make interactions feel mechanical.

• Resolution: AI should act as a peripheral support system (handling data synthesis, documentation, and subtle cues) rather than a central decision-maker, allowing the provider to remain the primary agent of human connection.

C. AI Opportunities and Concerns

• Opportunities: Providers expressed strong interest in AI for:
  • Extracting and prioritizing key patient data (e.g., "what is killing the patient?").
  • Synthesizing a "patient story" for rapid preparation.
  • Providing real-time, context-aware conversation prompts (e.g., suggesting a spiritual angle if the patient is religious).
  • Automating 80% of documentation to save time.
  • Providing immediate post-conversation feedback to improve skills.
• Concerns:
  • Loss of Authenticity: Fear that patients will perceive AI-assisted conversations as insincere.
  • Distraction: Technology must not interrupt eye contact or the flow of conversation.
  • Privacy/Consent: Strict adherence to consent for ambient listening.

5. Design Guidelines (DG)

Based on the findings, the authors propose four design guidelines for AI systems in EDs:

• DG1: At-a-Glance Information Scaffolding.

  • AI should automatically extract, synthesize, and summarize key data (code status, goals, mortality risk) into a readable format.
  • It must highlight gaps (e.g., "No prior SIC found") to prevent wasted time searching for non-existent records.
  • It should offer context-specific entry points to start the conversation smoothly.

• DG2: Non-Intrusive Ambient and Peripheral Support.

  • During the conversation, AI should operate on peripheral displays (secondary screens) or via subtle cues, never interrupting the provider-patient eye contact.
  • It should act as a "listener" that flags emotional cues or suggests questions only during natural pauses.
  • It must avoid scoring or judging emotions to prevent surveillance feelings.

• DG3: Rationale-Driven and Provider-Curated Documentation.

  • AI should generate drafts that capture the rationale behind decisions (the "why"), not just the outcome.
  • The system must allow providers to easily edit, verify, and approve the notes, ensuring clinical autonomy.
  • It should track the longitudinal progress of SICs across different departments to ensure continuity.

• DG4: Immediate Post-SIC Feedback.

  • To break the "anxiety loop" and improve skills, AI should provide immediate, constructive feedback after the conversation.
  • Feedback should highlight strengths (to build confidence) and offer specific, actionable phrasing alternatives for missed opportunities, acting as a "reflection-on-action" tool.

6. Significance

This paper is significant because it moves beyond technical feasibility to address the socio-technical complexities of AI in high-stakes healthcare.

• Human-Centered Design: It redefines the role of AI not as an autonomous agent but as a tool that sustains the conditions necessary for human empathy.
• Workflow Integration: It provides a concrete, empirically grounded framework for integrating AI into the chaotic, time-constrained reality of the ED.
• Bridging the Gap: It addresses the critical gap between the need for early SICs and the systemic barriers preventing them, offering a pathway to reduce unnecessary suffering and improve care alignment for vulnerable populations.
• Future Research: It sets the stage for participatory design and prototyping of ambient AI systems that respect the "emotional labor" of healthcare providers.