Hospital acquired pressure injuries (HAPI) are a common and serious problem among surgical critical care patients. Although pressure injuries are common, some pressure injuries (PIs) can be prevented using measures such as specialty beds that are not feasible for every patient due to cost1. However, decisions about which patient would benefit most from a specialty bed are problematic because existing pressure injury risk identification tools identify nearly all critical care patients as ‘high risk’2.
The purpose of our study was to use a machine learning approach to develop a pressure injury model that can be used to allow clinicians to differentiate among critical care patients and identify those at highest risk.
We used patient data from the electronic health record to develop a model to predict pressure injuries among surgical critical care patients at a level 1 trauma center. We performed our analysis in R version 3.3.2 via the RStudio interface. We utilized multiple imputation for missing data. We divided our data into training (67%) and testing (33%) datasets and then developed a random forest (RF) to predict category 2 and greater PI development among patients in the training data set and finally applied it to the test data set. We used the testing data set to fit the receiver operating characteristic (ROC) curve. Predictor variables were delirium, hypotension, level of consciousness, oxygenation, sedation, severity of illness, fever, vasopressor infusions, body mass index, albumin, creatinine, glucose, hemoglobin, lactate, pre-albumin, and surgical time.
The final sample consisted of 6,376 patients. Two hundred and eighty-three individuals (4.4%) developed HAPIs of category 2 or greater. The area under the ROC Curve was 0.79. Although all the variables were included in the random forest, the most important according to the mean decrease in accuracy demonstrated a significant scree and were: body mass index, surgery time, creatinine, hemoglobin, and age.
Our model is qualitatively different than the other models in the literature because ours does not require clinicians to input a score into a decision-making tool (such as the Braden scale, Norton scale, etc.). To our knowledge, ours is the only model that uses data that are readily available in the electronic health record. Next steps include validation in an independent sample, and then after demonstrating initial external validity of the overall classifier, calibration to optimize specificity so that the model can be used to identify patients who will benefit most from interventions that cannot be extended to all patients, such as costly specialty beds. The finding that surgical time was a relatively important variable in the analysis warrants further study as few studies have examined duration of surgery in relation to pressure injury risk.
Learning Objective 1: After participating in this session, the learner should be able to evaluate the performance of our machine learning model in predicting the pressure injury outcome among critical care patients.
Jenny Alderden (Presenter)
Boise State University
Mollie Cummins, University of Utah
Ginnette Pepper, University of Utah
Andrew Wilson, University of Utah