Introduction: Pressure injury (PI) rates are highest in the critically ill. Currently, nurses assess patient’s PI risk with the Braden scale; however, the scale does not include factors demonstrated in the research that impact ICU patient’s PI risk. Machine learning models incorporate large numbers of variables, including static (e.g. demographics and co-morbidities) and dynamic (e.g. vital signs and vasopressors) risk factors. The purpose of this study was to develop a PI prediction model to examine static and dynamic PI risk factors and predict new occurrences of hospital-acquired pressure injuries (HAPI) using electronic health record (EHR) data in critically ill adults. Methods: We used EHR data from 7 critical care units from September 2022-August 2023; patients with a new ICU HAPI occurrence, case cohort, and patients without a new HAPI occurrence, control cohort, were matched by length of ICU stay in a case-to-control ratio of 1:10. For prediction modeling, we extracted 72 hours of data as the observation window with a 12-hour lead-time prior to HAPI occurrence or controls’ matched timestamp in the ICU. For missing data, imputation was performed using patients’ own data, if available prior to the observation window. Variables with high collinearity were removed before model build. The top 20 features were selected using ANOVA F-values to prevent overfitting and improve generalization. Tree-based models were used to preserve feature explainability. We used an 80/20 training/test split and applied SMOTE resampling to the training set to balance the cohorts. Results: We identified 252 new HAPI occurrences as the case cohort and 2259 patients as the control cohort. The SHAP analysis identified several dynamic variables as the strongest predictors, such as SpO2, PCO2, albumin, hemoglobin, I/O Daily Net Volume, length of hospital stay, urethral catheter, endotracheal tube, surgical airway, or mechanical ventilation, vasopressors and paralytic medication use in OR. The random forest models with and without the Braden scale sub-scores achieved high AUROC results of 0.981 and 0.962, respectively. Conclusions: We developed random forest models to identify ICU patients at high risk of developing a HAPI. We plan to explore time series-based models to enable continuous risk monitoring for critically ill patients.
Schallom et al. (Sun,) studied this question.