A Weighted Sample Framework to Incorporate External Calculators for Risk Modeling

Personalized risk prediction calculators abound in medicine, and they carry important information about the effect of prognostic factors on outcomes of interest. How to use that information in order to analyze local datasets is a pressing question, and several recent proposals have attempted to pool information from external calculators to local datasets using parameter sharing approaches. Here, we adopt a weighting approach using convex optimization in order to transfer information. Rather than directly modeling parameters, we instead pool information on a per-sample basis. In particular, we develop prediction-guided analyses, along with an attendant inferential strategy, for incorporating information from the external risk calculator. We also supplement this analytical approach with an exploratory technique using trees to describe what we term as 'calculator-guided observations.' In addition, the optimization problem itself can yield insights on the potential transferability of the external calculator to the local dataset. The methodology is illustrated by simulation studies as well as an application of risk calculators to the prediction of sentinel lymph node positivity in melanoma.

Automated summary

This study introduces a weighting approach using convex optimization to transfer information and develop prediction-guided analyses by incorporating information from external risk calculators to local datasets. It also utilizes tree structures to describe 'calculator-guided observations' and sheds light on the potential transferability of external calculators to local datasets.