Predicting Admissions and Emergency Visits from Time-lagged ACO-specific Data

Author: Chess Stetson

Using historical claims data to predict adverse events could present huge benefits to patients and healthcare providers. Yet the data necessary to make accurate predictions may not be immediately available to those healthcare stakeholders that are in a position to act upon it. We address this problem for Accountable Care Organizations (ACOs), who receive an extract of patient claims at a delay of 60-90 days after initial recording. We show that even with a 90-day billing lag, our system can predict ED visits and hospital admissions at a high level of statistical significance (p<10^-10, Mann-Whitney U-test). Expectedly, the prediction strength is weaker than if the billing lag were shorter. The Area Under the Receiver Operating Curve (AUC) for predicting admissions is .65 with a 90-day lag, as compared to .75 with a 1-day lag, and for predicting ED visits, AUC is .58 at a 90-day lag vs .69 at a 1-day lag. We discuss how these predictions can be used to inform new and supplemental data collection, and triage patients into risk categories. We also discuss alternate learning techniques (Random Forest and Deep Belief Nets), as well as graph-based techniques in the Helynx software tool for reducing complexity in heterogeneous health data, and their affect on predictive performance.

Co Author/Co-Investigator Names/Professional Title: Chess Stetson, Ph.D., Boris Revechkis, PhD., Diane Pham, M.D., Warren Hosseinion, M.D.

Funding Acknowledgement (If Applicable): Apollo Medical Holdings Inc