Under newer payment schemes, hospitals are penalized when discharged patients shortly find themselves readmitted. More accurate patient readmission risk models would help patients and hospitals.
Bradley Malin, Ph.D., Daniel Fabbri, Ph.D., Lina Sulieman, M.S., and colleagues used machine learning and electronic health record data to develop readmission models that significantly outperform conventional models, achieving up to 90 percent accuracy when combining post discharge data with data from before and during admission.
Using data from Vanderbilt University Hospital, the team developed readmission models for hip fracture and heart failure. A conventional model used for comparison tracked length of stay, acuity, accompanying diseases and conditions and emergency room visits. This model, with all data gathered at discharge, proved little better than tossing a coin.
Additional data used by the team included time since discharge, lab test data and counts of medications, clinic appointments and patient messages, among other data.
The study appears in the American Medical Informatics Association Annual Symposium Proceedings.
The Vanderbilt team was joined by investigators from Cornell University and IBM. This research was supported by IBM and the National Institutes of Health (grant TR000445).
Send suggestions for articles to highlight in Aliquots and any other feedback about the column to aliquots@vanderbilt.edu
