Development of an Early Prediction Model for Patients with Pressure Injury: Based on Explainable Machine Learning Methods

Aim: This study aims to develop an interpretable machine learning model for the early prediction of hospital-acquired pressure injuries (PIs). Design: Retrospective cohort study. Methods: A retrospective study design was employed using electronic health record (EHR) data from hospitalized patients to predict pressure injuries (PIs). The dataset was randomly divided into a training set (70%) and an independent test set (30%). Five machine learning algorithms were developed and compared: eXtreme Gradient Boosting (XGBoost), Decision Tree (DT), Naive Bayes (NB), Multilayer Perceptron (MLP), and Random Forest (RF). SHAP (SHapley Additive exPlanations) was applied to interpret the best-performing model. Results: A total of 1524 patients were included, with a pressure injury (PI) incidence of 9.78% (149/1524). The XGBoost model achieved the highest predictive performance, yielding an area under the curve (AUC) of 0.910. In contrast, the Naive Bayes model demonstrated limited generalizability (AUC = 0.841) and relatively poor predictive accuracy. SHAP analysis identified the top 15 predictors of PIs, among which the use of sedative–analgesic drugs, albumin level, and prothrombin time were the most highly predictive. Conclusions: The study successfully developed a machine learning model that enhances the prediction of pressure injuries (PIs) in hospitalized patients. The model, combined with SHAP-derived interpretability, may facilitate early interventions and ultimately reduce the incidence of pressure injuries.