Statistical Methodology and Engineering for Next Generation Clinical Risk Calculators

In today’s practice of medicine, a variety of online clinical risk calculators are available to assist doctors and patients in informed decision-making. These tools may have unparalleled accuracy when founded on large cohorts or clinical trial populations; they may have passed the litmus test of multiple validations. However, evolving clinical practice, technology and population characteristics, as well as the discovery of new markers, can quickly outdate an existing risk tool, making it non-optimal for the contemporary patient. The traditional path of waiting for the next clinical trial or grant collective to end in order to amass fresh data and build a brand new model is too slow for today’s rapid science society, suggesting novel re-calibration methods applied to compartmentalized models that can be incrementally updated in real time. While Electronic Health Records promise an inexpensive, uninhibited and institution-tailored data flow, the percent usable data can be crippled by selection bias, non-ignorable missing data mechanisms and entanglement in indeterminate text fields, requiring novel big-data and record-linkage approaches to unravel. In this chapter we outline statistical methods and engineering approaches that can be used to tackle these challenges, and thereby keep risk calculators up to date in a continually evolving clinical care landscape. To illustrate we outline our experience adapting the Prostate Cancer Prevention Trial Risk Calculator during the past decade to meet the evolving challenges to risk prediction, and new research needed for the next generation of clinical risk prediction tools.