Machine Learning Yields Fresh Insights Into Pressure Injury Risks

Newswise — Big data and machine learning helped develop a model for predicting risk for pressure injuries in critical care patients, according to new research published in the November issue of American Journal of Critical Care (AJCC).

In “Predicting Pressure Injury in Critical Care Patients: A Machine-Learning Model,” the research team examined five years of data on patients admitted to the adult surgical or surgical cardiovascular intensive care units at the University of Utah Hospital in Salt Lake City. Among the sample of 6,376 patients, hospital-acquired pressure injuries of stage 1 or greater developed in 516 patients, and injuries of stage 2 or greater developed in 257 patients.

With these two outcome variables identified, the researchers used machine learning to effectively and efficiently look at the large amount of clinical data readily available in the patient records and examine the relationships among the available predictor variables. They used a technique called random forest, which is relatively unaffected by moderate correlations among variables, an important characteristic because correlations among clinical variables are common in health research.

Principal investigator Jenny Alderden, PhD, APRN, CCRN, CCNS, is an assistant professor in the School of Nursing, Boise State University, Boise, Idaho, and an adjunct assistant professor in the College of Nursing, University of Utah, Salt Lake City. In addition to Boise State and the University of Utah, other members of the research team represent the University of Washington, Seattle, and Rocky Mountain University of the Health Professions, Provo, Utah.

The researchers believe their study is the only one in which machine learning was used to predict development of pressure injuries in critical care patients.

“Current risk-assessment tools classify most critical are patients as high risk for developing pressure injuries and therefore do not provide a way to differentiate among critical care patients in terms of pressure injury risk,” Alderden said. “Eventually, our model may offer additional insight to clinicians as they develop a plan of care for patients at highest risk and identify those who would benefit most from interventions that are not financially feasible for every patient.”

Among the variables that were most important according to the model’s mean decrease in accuracy was time required for surgery, an element that has not been well studied as a potential contributor to risk for pressure injury. Body mass index, hemoglobin level, creatinine level and age were also ranked as important variables on the basis of the model’s mean decrease in accuracy. The mean decrease in accuracy, which reflects complex relationships among variables, is assessed by temporarily removing a variable from the analysis and evaluating the change in model performance.

Eventually, the model could help identify which patients are at the greatest risk for developing pressure injuries and who would benefit from interventions such as specialty beds or more frequent skin inspection. The next step will be to validate and evaluate the model in a new sample of patients.  

Alderden is also co-author of a related study published in the November issue of the journal. In “Outcomes Associated With Stage 1 Pressure Injuries: A Retrospective Cohort Study,” the researchers examined the outcomes of stage 1 hospital-acquired pressure injuries among 6,377 critical care patients at the University of Utah Hospital over a five-year period. Their review indicated that 259 patients experienced stage 1 pressure injuries, with the injuries becoming worse in about one-third of the patients. Their findings suggest that nurses should consider maximum treatment for patients with stage 1 pressure injuries, particularly for patients who are older or who experience alterations in oxygen delivery or perfusion.

The two studies were supported by the National Institute of Nursing Research, part of the National Institutes of Health.

The researchers’ complete machine-learning procedure is described in a special supplement to the journal, available only on the AJCC website. To access the articles and the full-text PDF, visit the AJCC website at www.ajcconline.org.

About the American Journal of Critical Care: The American Journal of Critical Care (AJCC), a bimonthly scientific journal published by the American Association of Critical-Care Nurses, provides leading-edge clinical research that focuses on evidence-based-practice applications. Established in 1992, the award-winning journal includes clinical and research studies, case reports, editorials and commentaries. AJCC enjoys a circulation of more than 120,000 acute and critical care nurses and can be accessed at www.ajcconline.org.

About the American Association of Critical-Care Nurses: Founded in 1969 and based in Aliso Viejo, California, the American Association of Critical-Care Nurses (AACN) is the largest specialty nursing organization in the world. AACN represents the interests of more than half a million acute and critical care nurses and includes more than 200 chapters in the United States. The organization’s vision is to create a healthcare system driven by the needs of patients and their families in which acute and critical care nurses make their optimal contribution.

American Association of Critical-Care Nurses, 101 Columbia, Aliso Viejo, CA 92656-4109;

949-362-2000; www.aacn.org; facebook.com/aacnface; twitter.com/aacnme