Theoretical Exploration of Error Thresholds for Clinical AI Decision Support in Nursing: Exploratory Simulation Study...

Area of Science:

• Nursing informatics
• Human-computer interaction
• Clinical decision support systems

Background:

• Clinical artificial intelligence (AI) decision support is emerging in nursing.
• Existing large language models (LLMs) show moderate accuracy on complex clinical tasks.
• The required AI accuracy for safe clinical use remains unclear, varying with clinician experience and task complexity.

Purpose of the Study:

• Develop an empirically calibrated simulation model of human-AI reliance and error in nursing.
• Estimate the AI accuracy needed to achieve specific clinical error rate targets.

Main Methods:

• A linear reliance model incorporating AI accuracy, clinician experience, and task complexity was calibrated.
• Weighted least squares regression used 9 data points from 3 randomized experiments (N=3502).
• Predicted error was calculated as reliance × (1-AI accuracy) across a 27-cell factorial design.

Main Results:

• The model found AI accuracy (βA) to be significantly predictive of error (0.201; P(βA>0)>.99).
• For novice clinicians and high-complexity tasks, AI accuracy of 0.89 and 0.78 is needed for <10% and <20% error rates, respectively.
• Current LLMs (0.5-0.7 accuracy) predict 26%-41% error rates in this high-risk scenario.

Conclusions:

• Achieving <10% error for novice nurses on complex tasks requires AI accuracy of ~0.89.
• Current general-purpose LLMs may not meet this threshold for complex clinical tasks.
• The 'A-threshold' framework offers a tool for evaluating minimum AI accuracy requirements, pending nursing-specific validation.