Gradient boosting for group testing

When conducting disease screening within a population, it is often more efficient and cost effective to group individual specimens and test them in pools, rather than testing each specimen individually. This method, which is known as group testing, speeds up the diagnostic process and reduces costs, especially when the outcome of interest is rare. However, the data collected from group testing can be inherently complex, especially in the presence of imperfect testing, and this complexity can hinder surveillance efforts. To overcome this challenge, we propose a gradient boosting framework designed to build predictive models based on group testing data using individual-level predictors. Our framework is flexible and supports a wide range of weak learners, including regression trees, kernel smoothing, and splines. In addition, our model accommodates data arising from any group testing protocol and accounts for the effects of imperfect testing. To optimize model performance, we develop a cross-validation approach that selects optimal tuning parameters for the weak learners. We explore the performance of our approach through numerical studies and illustrate our methods using chlamydia group testing data collected by the State Hygienic Laboratory in Iowa.