Predicting Pressure Injury in Critical Care Patients: A Machine-Learning Model

Abstract

<div class="section"> <a class="named-anchor" id="S1"> <!-- named anchor --> </a> <h5 class="section-title" id="d3002907e181">Background</h5> <p id="P1">Hospital-acquired pressure injuries are a serious problem among critical care patients. Some can be prevented by using measures such as specialty beds, which are not feasible for every patient because of costs. However, decisions about which patient would benefit most from a specialty bed are difficult because results of existing tools to determine risk for pressure injury indicate that most critical care patients are at high risk. </p> </div><div class="section"> <a class="named-anchor" id="S2"> <!-- named anchor --> </a> <h5 class="section-title" id="d3002907e186">Objective</h5> <p id="P2">To develop a model for predicting development of pressure injuries among surgical critical care patients. </p> </div><div class="section"> <a class="named-anchor" id="S3"> <!-- named anchor --> </a> <h5 class="section-title" id="d3002907e191">Methods</h5> <p id="P3">Data from electronic health records were divided into training (67%) and testing (33%) data sets, and a model was developed by using a random forest algorithm via the R package “randomforest.” </p> </div><div class="section"> <a class="named-anchor" id="S4"> <!-- named anchor --> </a> <h5 class="section-title" id="d3002907e196">Results</h5> <p id="P4">Among a sample of 6376 patients, hospital-acquired pressure injuries of stage 1 or greater (outcome variable 1) developed in 516 patients (8.1%) and injuries of stage 2 or greater (outcome variable 2) developed in 257 (4.0%). Random forest models were developed to predict stage 1 and greater and stage 2 and greater injuries by using the testing set to evaluate classifier performance. The area under the receiver operating characteristic curve for both models was 0.79. </p> </div><div class="section"> <a class="named-anchor" id="S5"> <!-- named anchor --> </a> <h5 class="section-title" id="d3002907e201">Conclusion</h5> <p id="P5">This machine-learning approach differs from other available models because it does not require clinicians to input information into a tool (eg, the Braden Scale). Rather, it uses information readily available in electronic health records. Next steps include testing in an independent sample and then calibration to optimize specificity. ( <i>American Journal of Critical Care</i>. 2018; 27:461–468) </p> </div>