B0330
Title: Globally-accessible and individual-tailored clinical risk prediction
Authors: Donna Ankerst - Technical University of Munich (Germany) [presenting]
Abstract: Six commonly used logistic regression methods for accommodating missing risk factor data from multiple heterogeneous cohorts, in which some cohorts do not collect some risk factors at all, were compared in order to develop an optimal flexible online prostate cancer risk prediction tool. All users had to have prostate-specific antigen and age, but the remaining ten risk factors were optional, yielding 1024 possible missing data patterns. The six methods included three variations of available case methods that fit models to data according to available risk factors from the user. The remaining three methods used all training data and included variations that explicitly modelled or imputed missing risk factors. Analysis of over 10,000 biopsies from ten North American and European cohorts for model training/internal validation, and over 5,000 biopsies for external validation yielded the available cases pooled main effects method as optimal. Developers of clinical risk prediction tools should optimize the use of available data and sources even in the presence of high amounts of missing data. For the end-user, developers should offer options for missing risk factors.