Targeted Recalibration Using Single-knot Splines to Improve the Clinical Utility of a Risk Model

Abstract

Risk prediction models are sometimes used to guide clinical decisions. Specifically, a patient may decide for or against an intervention based on whether his/her estimated risk falls above or below a critical threshold. When a risk model is used to make decisions in this way, it is especially important that the predicted risks are well calibrated. Miscalibrated risks are misleading and, moreover, reduce the clinical utility of risk-based decision-making to the patient population. Risk model miscalibration can occur due to overfitting or when a risk model is developed in one population and applied to another population. Recalibration methods can be used to address risk model miscalibration, but most methods do not account for how the risk model will be applied. We propose a new method of recalibration that is designed for risk models that will be used for risk-based decision-making. The method uses splines with a single, strategically placed knot to add flexibility and target good calibration where it is most important: at the critical threshold used for decision-making. We present simulation studies that compare our proposed recalibration method to existing methods and demonstrate our method's ability to improve the clinical utility of risk models in a variety of settings.