New AI Model May Improve Hospital Patient Care and Outcomes

Researchers use machine learning algorithms to predict pressure injury risk.

Posted Mar 01, 2021

Source: Geralt/Pixabay

Scientists are increasingly finding more ways to apply the predictive capabilities of artificial intelligence (AI) machine learning in health care, life sciences, and medical fields. In a recent study published in the Journal of the American Medical Informatics Association, researchers use AI machine learning and electronic health records to predict individuals at risk for pressure injuries with high precision.

When an area of skin is injured due to force on the skin’s surface, this is called a pressure injury. Other names for this condition are pressure sore, decubitus ulcer, and pressure ulcer according to the Cleveland Clinic. A more familiar name for a pressure injury is bedsore.

The cause of this type of injury may be due to constant pressure on the skin, usually over the bony parts of the body such as the tailbone, hips, elbows head, heels, and ankles. Another cause may be due to a shearing or frictional force between the skin and another surface, such as when a patient is moving in a wheelchair.

Each year, there are 2.5 million patients affected by pressure injuries resulting in roughly 60,000 deaths, over 17,000 lawsuits, and costs ranging from USD 9.1 to 11.6 billion according to statistics from the Agency for Healthcare Research and Quality at the U.S. Department of Health and Human Services. Caring for pressure injuries is expensive. The cost per pressure ulcer for one patient can range anywhere from USD 20,900 to up to USD 151,700 according to the agency.  

“Pressure injuries contribute adversely to an individual’s morbidity, mortality, and physical and psychosocial quality of life,” wrote researchers Wenyu Song, Min-Jeoung Kang, David Bates, and Patricia Dykes at the Brigham and Women’s Hospital in Boston, Massachusetts, in collaboration with Linying Zhang and Jiyoun Song at Columbia University, and Wonkyung Jung at the University of Washington. “They are also expensive. Pressure injuries increase hospital length of stay by 4-10 days and total healthcare costs by $10 708 per patient, accounting for approximately $26.8 billion per year.”

The researchers set out to create a predictive model using AI machine learning and phenotypes from electronic health record data that includes full assessments by nurses. The data used was from over 188,500 anonymized inpatient clinical records from five hospitals at Mass General Brigham spanning the period of 2015-2018.

The team created a list of features from scientific literature and reviewed it by clinical domain experts, then validated it against the Communicating Narrative Concerns Entered by RNs (CONCERN) database resulting in a cohort of 9,148 patients.

The researchers grouped the top 10 clinically significant features into three categories: neurological assessment for Glasgow coma scale and consciousness, physical mobility for gait/transferring, activity, and spinal cord injury, and blood chemistry panel for albumin, hemoglobin, blood urea nitrogen, chloride, and creatine.  

Then the researchers created four pressure injury prediction models using the features with logistic regression (LR), support vector machines (SVM), random forest (RF), and neural network algorithms. The neural network was implemented using the Keras library in Python, and the remaining three algorithms used the Scikit-Learn library in Python.

The researchers found that logistic regression was outperformed by the other three machine learning models and that the random forest model achieved the highest Area Under the Curve (AUC), a widely used metric for evaluation of the performance of an AI classification model. According to the researchers, their random forest machine learning model achieved 94 percent AUC based on a five-fold cross-validation.

“Our AUC was over 90% and may be used as a prediction tool in clinical practice and as a baseline model in future pressure injury studies,” concluded the researchers. “Our models derived from both hospital and nonhospital acquired pressure injury events could provide valuable information to clinicians and nurses to facilitate early prevention of these distinct types of pressure injury. The strong relationship between nurse-assessment features and occurrence of pressure injury revealed in our results could also help nurses to identify high-risk hospitalized patients and inform tailored preventative interventions.”

Copyright © 2021 Cami Rosso All rights reserved.