Hierarchical Latent Factor Models for Improving the Prediction of Surgical Complications Across Hospitals

• Activity Number: 460 - Clustering Methods for Big Data Problems
• Type: Topic Contributed
• Date/Time: Wednesday, August 2, 2017 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Elizabeth Lorenzi* and Katherine Heller and Ricardo Henao and Zhifei Sun
• Companies: Duke University and Duke University and Duke University and Duke University
• Keywords: transfer learning ; factor model ; bayesian nonparametrics ; healthcare ; machine learning ; covariance estimation

Abstract

We develop a framework for transfer learning, inspired by improving surgical predictions across hospitals, using hierarchical latent factor models. These factor models are used to learn the dependence structure between a larger source dataset (surgeries across all hospitals nationally) and a target dataset (surgeries at a local hospital). A hierarchical prior is put on the loadings matrix to appropriately account for the different covariance structure in our target and source populations. We extend this model to handle more complex relationships between the populations by deriving a stick-breaking formulation for the latent factors, allowing our model to flexibly vary the number of factors needed for each population. This work is motivated by our goal of building a risk-assessment model for surgery patients, using both institutional and national surgical outcomes data. The national surgical outcomes data is collected through NSQIP (National Surgery Quality Improvement Program), a database housing almost 4 million patients from over 700 different hospitals. We find that our methodology can model surgical complications well, improving overall predictions at Duke University Hospital.