What Lab Test Should I Perform Next for This Patient? Feature Acquisition of Subsets at Test-time With Tractable Models

We address a problem setting where given a history of examples with all features, the goal is to predict the best subset of features to acquire for a new example that only has baseline features. This problem is inspired by clinical settings, where some features such as demo- graphics are cheaply and easily obtained, while others such as blood tests and MRIs may be more costly, time-consuming, or invasive. We propose the Feature Acquisition of Subsets at Test-time (FAST) algorithm, which uses a tractable probabilistic model during training to efficiently compute the best subsets of past examples with all features (such as rigorously tested patients from a clinical study), so it can learn to use only baseline features to predict the single best subset of features to acquire for a new example during testing (such as using demographics to predict the best lab test for a new patient entering a clinic). Motivated by medical settings, we present the effectiveness of FAST on four real medical data sets.