|
Abstract:
|
In several disease areas, efforts are underway to combine biomarkers and clinical predictors to assist in disease screening, diagnosis, and prognosis. Under current strategies, novel markers that have high incremental value are sought with the goal of yielding a combination that may improve classification or prediction accuracy. However, little consideration is given to the costs of measuring those markers. That is, given the choice between two candidate markers, current methods would select the one that results in higher incremental performance with regards to some criteria, such as model fit or accuracy. However, incremental performance must also be considered with respect to the incremental cost of measuring the selected marker. A small performance increment may not be cost-effective, especially if improved performance has little impact on decision-making and subsequent clinical outcomes. To address this issue, we propose a statistical boosting algorithm for cost-effective model selection, so that model accuracy is optimized while accounting for marker measurement costs. We evaluate our proposed methods using simulation studies and illustrate them on a cystic fibrosis dataset.
|
