Technical Summary: CandorMD – An AI-Assisted Audio Simulation and Feedback System for Medical Error Disclosure

1. Problem Statement

Medical error disclosure is a high-stakes, ethically critical, yet emotionally complex communication task that clinicians frequently find difficult to navigate. Effective disclosure requires balancing transparency, empathy, and accountability while managing the clinician's own emotions and the patient's potential anger, grief, or distrust.

Current training methodologies suffer from significant limitations:

• Pedagogical Gaps: Reliance on didactic lectures and passive observation ("see one, do one, teach one") provides insufficient active practice.
• Scalability and Adaptability: Standardized patient encounters and role-playing are resource-intensive and difficult to scale across diverse specialties.
• Static Tools: Existing Video-based Communication Assessment (VCA) tools use pre-recorded scenarios that lack dynamic adaptability. They cannot respond to learner inputs in real-time, failing to mimic the reciprocal nature of actual conversations.
• Feedback Latency: Feedback is often delayed, generic, or reliant on human raters, preventing "just-in-time" learning and failing to address specific individual needs or emotional nuances.

These gaps contribute to clinician avoidance of disclosure conversations, reduced self-efficacy, and compromised patient trust.

2. Methodology

The research team developed CandorMD, an AI-assisted audio simulation and feedback prototype, and evaluated it through a mixed-methods approach involving design probes and semi-structured interviews.

2.1 System Architecture

CandorMD is built around two coordinated AI agents operating over a shared conversation state but maintaining distinct information access:

• Patient Simulation Agent:

  • Function: Generates realistic patient responses (text and audio) that adapt dynamically to the clinician's utterances.
  • Internal State: Maintains an Information Context (what the patient knows vs. suspects), an Affective State (continuous values for emotions like anger, grief, trust), and Conversation Memory.
  • Audio Generation: Unlike simple text-to-speech, this agent generates detailed voice instructions (pacing, emotional tone, prosody) alongside the text to ensure emotional delivery matches the content. It uses OpenAI's GPT-4o-mini-tts for audio synthesis.
  • Adaptivity: The patient's reaction shifts based on the clinician's performance (e.g., becoming more receptive if communication is effective, or withdrawing if it is not).

• Evaluator Agent:

  • Function: Provides feedback on the clinician's communication skills.
  • Feedback Modes:
    1. Turn-by-turn: Immediate, contextualized guidance during the simulation.
    2. Overall Summary: A post-session report highlighting strengths, areas for improvement, and specific suggestions.
  • Evaluation Framework: Based on established disclosure frameworks (e.g., SPIKES), the agent maps conversation stages to four feedback dimensions: Information Seeking, Emotional Expression, Trust & Accountability, and Forward-looking Care Planning.
  • Information Asymmetry Management: The Evaluator has access to the full clinical context (including the physician's knowledge), while the Patient Agent is restricted to what the patient realistically knows. To prevent "leaking" physician knowledge to the patient prematurely, the system filters the Evaluator's output into keywords before guiding the Patient Agent's next response.

2.2 Study Design and Participants

• Participants: 12 key informants recruited via professional networks and snowball sampling, representing four distinct roles:
  • Practicing Clinicians (4)
  • Risk Managers (2)
  • Communication Researchers (3)
  • Patient Advocates (3)
• Procedure:
  1. Semi-structured Interviews: Explored current practices, challenges, and the four Research Questions (RQs) regarding factors, practices, challenges, and technology-supported training.
  2. Prototype Interaction: Participants engaged with CandorMD using a think-aloud protocol, completing disclosure scenarios and providing in-situ verbalizations.
  3. Analysis: Thematic analysis was performed on interview transcripts and interaction logs to identify patterns in usability, simulation fidelity, and pedagogical value.

3. Key Contributions

The paper makes three primary contributions:

1. Empirical Insights: Detailed findings on the training needs and challenges in medical error disclosure, derived from interviews with a diverse ecosystem of stakeholders (clinicians, risk managers, advocates, and researchers).
2. System Design: The design and implementation of CandorMD as a "design probe" that addresses the limitations of current static training tools through multi-agent simulation, dynamic emotional modeling, and dual-mode feedback.
3. Design Recommendations: Empirically grounded recommendations for developing AI-supported communication training tools that extend beyond medicine, specifically regarding the balance of automation and human oversight, feedback timing, and emotional realism.

4. Results and Findings

4.1 Key Factors in Disclosure (RQ1)

• Trust: The foundation of disclosure; it must be built or repaired during the conversation, especially in acute care.
• Patient-Centeredness: Conversations must adapt to what the patient knows and their emotional state. It is a two-way interaction, not a monologue.
• Timeliness vs. Continuity: Disclosure should be prompt but is an ongoing process, not a single event.
• Transparency & Uncertainty: Clinicians must balance honesty with the admission of uncertainty, avoiding premature speculation.
• Apology: Viewed as an essential expression of empathy and accountability, distinct from a premature admission of legal fault.

4.2 Current Practices and Gaps (RQ2 & RQ3)

• Fragmented Training: Preparation is often ad-hoc or relies on observation. Structured models exist but are inconsistently applied.
• Emotional Regulation: Clinicians struggle with their own shame and guilt, which hinders communication.
• Feedback Deficits: Existing feedback is often delayed, inconsistent, or non-existent. Clinicians rely on patient reactions as the only feedback, which is costly and unstructured.
• Cultural Barriers: A culture that stigmatizes error or views communication skills as innate rather than learnable hinders improvement.

4.3 Prototype Evaluation (RQ4)

• Simulation Fidelity: Participants found the emotional trajectories (e.g., anger escalation) and handling of factual uncertainty realistic. However, some noted that prosody occasionally felt "artificial" or too formal.
• Feedback Quality: Participants valued the combination of immediate turn-by-turn guidance and end-of-session summaries. They preferred concrete exemplars ("You could say...") over generic rubric comments.
• Usability: The interface was generally intuitive. Suggestions included better differentiation between "stopping" and "ending" a case, and options to restart or course-correct mid-scenario.
• Use Cases: Participants envisioned two primary applications:
  1. Just-in-time rehearsal: Preparing for specific, upcoming difficult conversations.
  2. Curricular integration: Use in medical schools and residency programs for longitudinal skill building.

5. Significance and Claims

The paper positions CandorMD not as a final solution, but as a significant step toward addressing the "practice gap" in medical communication training.

• Enhancing Realism: By using multi-agent coordination to manage information asymmetry and dynamic emotional states, the system offers a more authentic simulation environment than static video tools.
• Scalability: The system provides a scalable alternative to resource-intensive standardized patient encounters, offering immediate, personalized feedback.
• Complementary Role: The authors explicitly state that AI cannot resolve systemic issues alone. The significance of the work lies in demonstrating how AI can complement traditional methods and address specific pedagogical gaps (real-time practice, emotional attunement) while acknowledging that structural, cultural, and policy reforms remain necessary for effective error disclosure.
• Generalizability: The design principles (multi-agent simulation, dynamic feedback, managing information asymmetry) are presented as transferable to other high-stakes healthcare communication scenarios beyond error disclosure.

The paper concludes that while technical solutions like CandorMD can enhance the realism and scalability of training, lasting improvement requires integrating these tools into institutional structures supported by mentorship and cultural change initiatives.