How AI helps health systems predict in-hospital mortality
AI models can predict when patients may be at an increased risk of in-hospital mortality, according to new research published in JAMA Network Open. If implemented, such models could be used to help healthcare providers improve decision-making and deliver better patient care.
“Approximately 2% of patients admitted to U.S. hospitals die during the inpatient admission,” wrote lead author Nathan Brajer, BS, Duke University School of Medicine in Durham, North Carolina, and colleagues. “Efforts to reduce preventable in-hospital mortality have focused on improving treatments and care delivery, and efforts to reduce nonpreventable mortality have focused on supporting patient preferences to die at home and attempting to reduce health care costs in the inpatient setting.”
Brajer et al. trained and validated a machine learning-based AI model with data from more than 31,000 adult patients who received care at a single hospital (hospital A) from Oct 2014 to December 2015. The model was then validated again with a separate dataset from hospital A and data from two additional hospitals (hospital B, hospital C).
“To make the model implementable at a system level, the model was trained on all adult patients using data elements commonly available across sites” they authors wrote. “The model only uses information from the current encounter without prehospital information, meaning that model outputs are accurate for patients who present for the first time to a health care setting. “
The team achieved in-hospital mortality rates of 3.0% for the initial hospital A dataset, 2.7% for the second hospital A dataset, 1.8% for hospital B and 2.1% for hospital C. Area under the ROC curve (AUROC) for those datasets ranged from 0.84 to 0.89.
An important detail of this study was that the researchers also evaluated their AI model prospectively, integrating it into an electronic health record.
“Prospectively evaluating the performance of machine learning models run on real-world, operational data is a crucial step toward integrating these models into the clinical setting and evaluating the impact on patient care,” the authors wrote.
The prospective dataset included more than 4,500 patients admitted from Feb. 14 to April 15, 2019. In this instance, the model’s in-hospital mortality rate was 1.6%, and its AUROC was 0.86.
“Taken together, the findings in this study provide encouraging support that machine learning models to predict in-hospital mortality can be implemented on live EHR data with prospective performance matching performance seen in retrospective evaluations of highly curated research data sets,” the team concluded.
Michael has more than 18 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.