To explore how AI explanations can assist primary eye care providers in evaluating AI predictions for glaucoma referrals and enhancing referral accuracy.
Key Findings:
Human-AI teams achieved 60% accuracy compared to 51% for humans alone.
Explanations did not enhance performance and introduced uncertainty about trusting AI.
Post-hoc explanations led to over-reliance on incorrect AI recommendations.
Both explanation types contributed to anchoring bias, aligning participants more closely with AI referrals.
Human-AI teams underperformed compared to AI alone, which had 80% accuracy.
Interpretation:
The study highlights the challenges of integrating AI in clinical decision-making, particularly regarding trust and reliance on AI recommendations among primary care providers, which can hinder effective patient care.
Limitations:
Explanations did not improve decision-making and may have worsened reliance on incorrect AI outputs, limiting their utility.
The study's sample size and context may limit generalizability, suggesting caution in applying findings broadly.
Conclusion:
Effective support mechanisms, such as training and user-friendly interfaces, are needed to enhance human-AI collaboration in glaucoma referrals while maintaining human agency.
