“If I Had All the Time in the World”: Ophthalmologists' Perceptions of Anchoring Bias Mitigation in Clinical AI Support

Title of the Paper

“If I Had All the Time in the World”: Ophthalmologists’ Perceptions of Anchoring Bias Mitigation in Clinical AI Support

Paper Information

• Subject Area: Research on cognitive bias and AI clinical decision support in ophthalmology
• Keywords: cognitive bias, anchoring bias, decision support, artificial intelligence, bias mitigation, clinical decision support systems (CDSS), human-computer interaction, ophthalmology, diabetic retinopathy (DR)

Research Background and Problem

• Problem or Challenge:

  • Artificial intelligence (AI) is increasingly applied in clinical decision support systems (CDSS) to assist physicians in handling complex decisions and reducing workload, but it may also introduce new cognitive biases (e.g., anchoring bias).
  • Specific biases like anchoring bias (over-reliance on initial information) can lead to diagnostic errors, adversely affecting patient health.
  • Misunderstandings of AI capabilities and limitations by medical professionals may exacerbate negative impacts.

• Significance:

  • With the rising prevalence of diabetic retinopathy (DR) cases, clinicians must frequently manage complex cases while ensuring diagnostic accuracy and efficiency.
  • How bias mitigation methods can be integrated into CDSS remains underexplored, particularly regarding usability and effectiveness in real clinical scenarios.

• Research Motivation and Related Work:

  • Existing studies indicate the widespread presence of cognitive biases, but there is limited exploration of how to mitigate these biases in AI-supported CDSS.
  • The authors aim to bring findings from bias mitigation literature into the ophthalmology context, studying clinicians’ perceptions of specific strategies and designing solutions based on their real workflows.

Solution

• Method or Solution:

  • The authors designed an interactive prototype integrating three bias mitigation strategies:
    1. “Hear the Story First”: Encourages clinicians to independently evaluate data before viewing AI-recommended results.
    2. “Decision Justification”: Requires clinicians to document their diagnostic reasoning to prompt reflection.
    3. “Consider the Opposite”: Prompts clinicians to reassess when no abnormalities are detected in their diagnosis but flagged by AI.

• Innovative Aspects:

  • Bias mitigation strategies were designed based on clinicians’ actual workflows.
  • Complex bias mitigation strategies were seamlessly integrated into the CDSS interface, providing intuitive tools to avoid cognitive biases.
  • Investigated clinicians’ perspectives on cognitive bias issues in AI decision support.

• Implementation Steps and Key Techniques:

  • Study 1: Conducted seven interviews and three field observations to collect contextual data on clinicians’ use of CDSS.
  • Study 2: Built a prototype and tested the effectiveness and acceptability of bias mitigation strategies with six clinicians.
  • Key Techniques: Combined user interface research with qualitative analysis (using reflexive thematic analysis) to analyze results.

Research Findings

• Specific Findings:

  • Clinicians provided varied feedback on the three bias mitigation strategies:
    • "Hear the Story First": Reduced anchoring bias but decreased efficiency.
    • "Decision Justification": Potentially improved diagnostic accuracy but required additional time, with some clinicians finding it redundant.
    • "Consider the Opposite": Helped avoid overlooking critical details, though some clinicians disliked the feeling of being “checked” by AI.
  • Acceptance of bias mitigation strategies varied among clinicians, influenced by their trust in current AI capabilities.

• Comparison with Existing Solutions:

  • The study highlighted real-world issues with current CDSS, such as low accuracy and limited functionality, failing to meet clinicians’ expectations.
  • Suggested designing more sensitive user interfaces and periodically activating bias mitigation techniques to alleviate concerns about efficiency loss.

• Experimental or Evaluation Results:

  • Bias mitigation strategies may improve diagnostic standards but pose new challenges (e.g., efficiency concerns).
  • The learning potential of bias mitigation techniques is particularly significant for younger or less experienced clinicians, while senior clinicians may be less receptive.

• Limitations and Future Directions:

  • Limitations include the lack of independent evaluation of individual bias mitigation strategies and specific implementations, and the study’s focus on ophthalmology.
  • Recommendations for future work:
    1. Explore more flexible bias mitigation interface designs.
    2. Test bias mitigation strategies in broader medical contexts.
    3. Further quantify the long-term effectiveness of bias mitigation strategies in practice.