Decisions are not all equal-Introducing a utility metric based on case-wise raters' perceptions

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Evaluating AI-based decision support systems (AI-DSS) is crucial but current metrics often neglect contextual information.
• Existing metrics may not fully capture the nuances of human-AI interaction and decision-making processes.

Purpose of the Study:

• Introduce a novel utility metric, weighted Utility (wU), for AI-DSS evaluation.
• Demonstrate the application of wU for both AI model evaluation and optimization.
• Highlight the importance of human-centered evaluation in critical domains.

Main Methods:

• Developed the weighted Utility (wU) metric based on rater perceptions of annotation hesitation and training case relevance.
• Compared wU with existing metrics like Net Benefit and error-based metrics.
• Applied wU in three realistic case studies for model evaluation and optimization.

Main Results:

• The wU metric generalizes existing evaluation metrics.
• wU provides a more flexible tool for AI model evaluation.
• Optimization using wU significantly improved model performance (AUC 0.862 vs 0.895, p<0.05), particularly for complex cases (AUC 0.85 vs 0.92, p<0.05).

Conclusions:

• Utility should be a primary concern in evaluating and optimizing machine learning models in critical fields like medicine.
• A human-centered approach is vital for assessing AI's impact on human decision-making.
• Information gathered during ground-truthing can enhance AI model evaluation.