Predictive case control designs for modification learning

TL;DR

This paper introduces predictive case control (PCC) sampling designs to efficiently gather outcome labels for model modification learning, reducing sample size and cost compared to simple random sampling.

Contribution

It proposes a novel PCC sampling method based on model scores, with mathematical justification and a computational framework for design evaluation.

Findings

PCC sampling achieves effective model modification with smaller samples.

Simulation shows PCC outperforms simple random sampling in information efficiency.

Application to radiology data demonstrates practical utility of PCC in clinical settings.

Abstract

Prediction models for clinical outcomes may be developed using a source dataset and additionally applied to new settings. Towards model external validation and model updating in the new setting, one procedure is model modification learning that involves the dual goals of recalibrating overall predictions as well as revising individual feature effects. Modification learning generally requires the collection of an adequate sample of true outcome labels from the new setting, which is frequently an expensive and time-consuming process, as it involves abstraction by human clinical experts. To reduce the abstraction burden for such new data collection, we propose a class of designs based on original model scores and their associated outcome predictions. We provide mathematical justification that the general predictive score sampling class results in valid samples for analysis. Then, we focus attention specifically on a stratified sampling procedure that we call predictive case control (PCC) sampling, which allows the dual modification learning goals to be achieved at a smaller sample size compared to simple random sampling (SRS). PCC sampling intentionally over-represents subjects with informative scores, where we suggest using the D-optimality and Binary Entropy information functions to summarize sample information. For design evaluation within the PCC class, we provide a computational framework to estimate and visualize empirical response surfaces of the proposed information functions. We demonstrate the benefit of using PCC designs for modification learning, relative to SRS, through Monte Carlo simulation. Finally, using radiology report data from the Lumbar Imaging with Reporting of Epidemiology (LIRE) study, we illustrate the application of PCC for new outcome label abstraction and subsequent modification learning across imaging modalities.