Among hospitalized adults, delirium is common, and is associated with increased mortality, cognitive decline and new-onset dementia. More than two-thirds of delirium cases are considered preventable.
To gauge risk of new-onset delirium in hospitalized adults, Siru Liu, PhD, Adam Wright, PhD, and colleagues used machine learning based on electronic health records of 34,035 intensive care patients who had received periodic assessments for delirium during their hospital stays.
As reported in the Journal of the American Medical Informatics Association, the team tested several algorithms, with a six-hour predictive model figuring as the all-around top performer. Combining so-called long short-term memory with a gradient-boosting technique, the model predicted new-onset delirium with 82% sensitivity and had an area under the receiver operating characteristic curve (ROC) of 95% (with 90% or greater considered outstanding, ROC gauges a predictive model’s overall clinical usefulness and can be interpreted as measuring how well a model ranks patients according to risk).
Others on the study include Joseph Schlesinger, II, MD, Allison McCoy, PhD, Thomas Reese, PharmD, PhD, Bryan Steitz, PhD, Elise Russo, MPH, and Brian Koh. The study was supported by the National Institutes of Health (AG062499, LM014097).
