A Method to estimate the confidence interval around a multivariable risk score prediction

Multivariable risk scores predict an individual patient’s risk of disease based on that person’s unique combination of risk factor values. When calculating a risk score such as Framingham Risk Score or Reynolds risk scores, one plugs an individual patient’s data into the model and gets a natural logarithm of the odds of disease as a point estimate. This can then be easily transformed into a probability estimate. However, despite the variable in the model has both a point estimate of beta coefficient and a standard error of the beta coefficient, the risk score produces only a point estimate of risk for each person, with no estimate of the uncertainty or precision of that risk estimate. There is currently no method to estimate the confidence interval around the risk estimate for a multivariable risk prediction model. We propose a resampling-based approach to simultaneously provide a risk probability of the risk as well as the precision of the risk probability for better understanding of the risk prediction.