Automatic Stratification of Tabular Health Data

Stratifying an outcome of interest across sub-groups is a ubiquitous technique for better understanding tabular data. This work efficiently scales stratification across multiple features simultaneously to identify the strata with the most unexpectedly high (or low) outcomes. We identified an anomalous sub-group of neonatal mortality outcomes in a large global health study. Scanning over subsets of data is an alternative to fitting regression models or interpreting machine learning prediction models.