Vanderbilt Health researchers used data from roughly 60,000 Tennessee surgical cases to train risk models for postoperative acute kidney injury (AKI), a common surgical outcome associated with increased mortality, increased hospital length of stay and readmission, and increased health care costs. A solution for efficiently ranking preoperative patients according to AKI risk could help advance prevention.
Matt Shotwell, PhD, Associate Professor of Biostatistics and Anesthesiology, Joel Bradley III, MD, Assistant Professor of Surgery in the Division of General Surgery, and colleagues undertook supervised machine learning on routine preoperative data captured in a national surgical registry. The team’s winning model achieved areas under the receiver operating characteristic curve (AUC) of 87% and 88% across validation sets from Tennessee and from the nation, respectively; AUC, with scores ranging from 50%-100%, gauges how well a model ranks patients according to risk. The team reported the study in the Journal of the American College of Surgeons.
Throughout, the study used data from the American College of Surgeons National Surgical Quality Improvement Program, which samples data from hospitals in the U.S. and Canada. The study’s external validation set included nearly 1 million cases.
The team’s winning machine learning-derived model is an additive logistic regression and thus readily interpretable — no black box here: The chief predictors in the model included inpatient status, excess fluid in the peritoneal cavity (ascites), renal failure, preoperative creatinine, sepsis, and physical-status classification (per the Americal Society of Anesthesiologists’ six-point grading system).
Others on the study from Vanderbilt Health included Cassandra Hennessy, MS, and Barbara Martin, MBA, RN. The team was joined by Stephen Behrman, MD, from Baptist University College of Medicine in Memphis.
