Who is Trusted for a Second Opinion? Comparing Collective Advice from a Medical AI and Physicians in Biopsy Decisions After Mammography Screening

Research Background and Problem

• What issues or challenges did the authors identify?
  As artificial intelligence (AI) is increasingly adopted in the medical field, its impact on patient decision-making remains unclear, especially when AI and doctors provide conflicting recommendations. Existing studies lack an understanding of how patients make decisions when faced with consistent or conflicting advice.

• Why is this issue important?
  In high-risk medical scenarios such as breast cancer screening, early and accurate diagnosis is critical for patient survival rates. Understanding patients' trust and adherence to collective recommendations from AI and doctors can help optimize medical workflows and design more effective AI decision-support systems.

• Research Motivation and Related Work
  The study aims to explore how patients perceive integrated medical decisions based on AI and doctors, particularly when the two recommendations conflict. The background includes the rapid application of medical AI in breast cancer screening and other diagnostic fields, where the technology enhances diagnostic accuracy but patients' trust in AI remains lower than in doctors. Cognitive Dissonance Theory is introduced to understand the psychological impact of conflicting recommendations on decision-making.

Solution

• What methods or solutions did the authors propose?
  The authors designed a mixed-method approach, including a qualitative interview study (9 patients) and a quantitative online experiment (339 participants), to evaluate patients' trust attitudes and behaviors in scenarios where AI and doctors jointly provide recommendations.

• What are the innovative aspects of the solution?

  1. In-depth analysis of how patients handle consistency and inconsistency in collective medical recommendations.
  2. Qualitative research reveals patients' unique expectations of AI (e.g., "four-eyes principle").
  3. Quantitative experiments use real-world case descriptions and decision-making data to explore the connection between patients' actual decisions and trust.

• What are the implementation steps and key techniques used?

  1. Qualitative Interviews: Gather insights from breast cancer patients regarding their views on AI diagnostic support, extracting key themes such as patient autonomy and AI's lack of empathy.
  2. Quantitative Experiment:
     • Use case scenarios and visual stimuli related to breast cancer screening to simulate real-life situations where patients face AI and doctor recommendations.
     • Cross-design different recommendation combinations (AI and doctor agreement or conflict).
     • Collect participants' trust attitudes (7-point Likert scale) and behavioral trust (adherence to recommendations) data.
  3. Data Analysis: Compare differences in trust and adherence between AI and doctors using t-tests and logistic regression modeling.

Research Findings

• What specific findings were achieved?

  1. Qualitative Research: Patients generally hope to achieve more rigorous decision-making through AI's second opinion ("four-eyes principle") while expressing concerns about AI's lack of care for individual and emotional needs.
  2. Quantitative Research:
     • Patients trust doctors more (average trust score of 5.01) than AI (average trust score of 3.71).
     • Under consistent recommendations, 99% of participants adhered to the advice to undergo a biopsy. However, when recommendations conflicted, participants exhibited risk-averse behavior, with 66% adhering to AI's advice (if AI recommended a biopsy).

• What advantages does this solution have compared to existing ones?
  Compared to studies that examine human doctors or AI in isolation, this research systematically analyzes patients' trust and decision-making behavior in collective medical recommendation models, uncovering the subtle relationship between risk aversion and trust dynamics.

• What are the experimental or evaluation results?

  • In conflict scenarios, participants significantly tended to make more conservative, risk-averse choices (e.g., undergoing a biopsy).
  • Even though trust in AI is lower than in doctors, when AI recommended a biopsy, 66% of participants adhered to AI's advice, indicating that risk perception may outweigh trust attitudes.

• Limitations and Future Directions:

  • Limitations:
    1. Data is based solely on German female participants, lacking geographic and gender diversity.
    2. The experiment uses virtual scenario simulations, which may differ from decision-making behavior under real-world high-pressure conditions.
  • Future Directions:
    1. Conduct cross-cultural studies on a global scale to validate the generalizability of the findings.
    2. Further explore emotional support and enhanced explainability features in medical AI design.
    3. Investigate trust and adherence patterns in more clinical scenarios (e.g., cardiology) involving AI-doctor collaboration.

The analysis above demonstrates how the study uses a mixed-method approach to deeply explore the trust dynamics of AI and doctors as joint recommendation providers in medical settings, integrating the complexity of trust, risk perception, and adherence behaviors. This provides significant insights for optimizing AI applications in healthcare in the future.