Heterogeneity between validation and training data used to build clinical risk prediction models biases the performance of the models on validation samples. A framework is developed to accommodate selection bias coming from heterogeneous distributions of risk factors and verification bias coming from different verification mechanisms between training and validation cohorts. Adjustments result in weighted versions of the usual performance metrics with different weights addressing different types of bias.     «

Heterogeneity between validation and training data used to build clinical risk prediction models biases the performance of the models on validation samples. A framework is developed to accommodate selection bias coming from heterogeneous distributions of risk factors and verification bias coming from different verification mechanisms between training and validation cohorts. Adjustments result in weighted versions of the usual performance metrics with different weights addressing different types...     »