How Can Clinicians Decide if AI-Enabled Risk Prediction Models Are Fit for Purpose for Real World Use?

Area of Science:

• Clinical informatics
• Medical artificial intelligence
• Patient safety

Background:

• Artificial intelligence (AI) models are increasingly used in hospitals for adverse event prediction.
• The beneficial impact of AI models varies due to issues with predictive accuracy, false alerts, and narrow prediction windows.
• Existing performance measures do not adequately capture the interplay between AI model effectiveness, efficiency, and utility.

Purpose of the Study:

• To present methods for assessing and optimizing AI risk prediction models in clinical practice.
• To address the limitations of current performance measures for AI models.
• To improve the clinical utility of AI-driven alerts for adverse events, using sepsis as a case study.

Main Methods:

• Discussing the interdependency of predictive accuracy, false alert burden, and prediction window length.
• Presenting methods to evaluate AI model suitability and determine optimal decision thresholds.
• Outlining strategies for optimizing AI model performance and configuring alert delivery systems.

Main Results:

• The commentary proposes a framework for a more holistic evaluation of AI risk prediction models.
• Methods are suggested to enhance the practical usability of AI tools for frontline clinicians.
• The approach aims to improve the effectiveness, efficiency, and utility of AI in preventing adverse patient events.

Conclusions:

• Clinicians need better tools to assess and utilize AI risk prediction models effectively.
• Optimizing AI models requires considering their accuracy, efficiency, and utility in real-world clinical settings.
• Implementing improved AI model evaluation and alert systems can enhance patient safety and hospital practice.