Learning decision thresholds for risk stratification models from aggregate clinician behavior.

Area of Science:

• Clinical Decision Support
• Health Informatics
• Predictive Modeling

Background:

• Risk stratification models are crucial for clinical practice, but selecting decision thresholds can be challenging.
• These thresholds determine when to trigger actions like electronic alerts, impacting patient care.
• Existing methods for threshold selection may not reflect real-world clinical behavior.

Purpose of the Study:

• To propose a flexible approach for learning decision thresholds from real-world physician practice.
• To demonstrate the feasibility of using real-world data to establish reference decision thresholds.
• To enable evaluation of clinical decision support tools against community standards of care.

Main Methods:

• Leveraging collective treatment decisions from real-life data to learn physician practice patterns.
• Utilizing a specific example: prescribing statins for cardiovascular disease primary prevention.
• Applying the 2013 pooled cohort equations to calculate 10-year cardiovascular risk.

Main Results:

• Demonstrated the feasibility of learning implicit decision thresholds from real-world data.
• Showcased how learned thresholds reflect existing physician behavior and community standards.
• Validated the approach for evaluating proposed operating points in clinical decision support.

Conclusions:

• Real-world data offers a viable method for learning and setting decision thresholds in clinical practice.
• This approach facilitates the auditing and monitoring of model-guided clinical decision-making.
• Aligning decision support tools with actual physician practice enhances their clinical utility.