Structured variable selection: an application in identifying predictors of major bleeding among hospitalized hypertensive patients using oral anticoagulants for atrial fibrillation

Predictor identification is important in medical research as it can help clinicians to have a better understanding of disease epidemiology and identify patients at higher risk of an outcome. Variable selection is often used to reduce the dimensionality of a prediction model. When conducting variable selection, it is often beneficial to take selection dependencies into account. Selection dependencies can help to improve model interpretability, increase the chance of recovering the true model, and augment the prediction accuracy of the resulting model. The latent overlapping group lasso can achieve the goal of incorporating some types of selection dependencies into variable selection by assigning coefficients to different groups of penalties. However, when the selection dependencies are complex, there is no roadmap for how to specify the groups of penalties. Wang et al. (2021) proposed a general framework for structured variable selection, and provided a condition to verify whether a penalty grouping respects a set of selection dependencies. Based on this previous work, we construct roadmaps to derive the grouping specification for some common selection dependencies and apply them to the problem of constructing a prediction model for major bleeding among hypertensive patients recently hospitalized for atrial fibrillation and then prescribed oral anticoagulants. In the application, we consider a proxy of adherence to anticoagulant medication and its interaction with dose and oral anticoagulants type, respectively. We also consider drug-drug interactions. Our method allows for algorithmic identification of the grouping specification even under the resulting complex selection dependencies.