Prediction impact curve is a new measure integrating intervention effects in the evaluation of risk models

Abstract

Objective: We propose a new measure of assessing the performance of risk models, the area under the prediction impact curve (auPIC), which quantifies the performance of risk models in terms of their average health impact in the population.

Study design and setting: Using simulated data, we explain how the prediction impact curve (PIC) estimates the percentage of events prevented when a risk model is used to assign high-risk individuals to an intervention. We apply the PIC to the Atherosclerosis Risk in Communities (ARIC) Study to illustrate its application toward prevention of coronary heart disease.

Results: We estimated that if the ARIC cohort received statins at baseline, 5% of events would be prevented when the risk model was evaluated at a cutoff threshold of 20% predicted risk compared to 1% when individuals were assigned to the intervention without the use of a model. By calculating the auPIC, we estimated that an average of 15% of events would be prevented when considering performance across the entire interval.

Conclusion: We conclude that the PIC is a clinically meaningful measure for quantifying the expected health impact of risk models that supplements existing measures of model performance.

Keywords: AUC; Coronary heart disease; Prediction impact curve; Predictive ability; Predictive model; Risk model.