Predicting Risk Factors of Pressure Injury for Perioperative Patients Through Machine Learning With SHapley Additive...

Perioperative pressure injuries (PIs) are a common complication of surgical hospitalization that can lead to serious health problems, compromise wound healing, and cause infections. This study aimed to predict the early risk of PI in perioperative patients by developing an interpretable machine-learning model.

Clinical data of perioperative surgical inpatients in a tertiary hospital in Shanghai from 2017 to 2023 were retrospectively analyzed, covering a total of 14,416 patients for screening. Feature variables were selected using the XGBoost model. Seven machine-learning (ML) models were employed to predict these variables. To evaluate the performance of these prediction models, various metrics such as accuracy, recall, and area under the curve were employed. In addition, a 10-fold cross-validation method was implemented to ensure the generalizability of the models. Furthermore, this study incorporated SHapley Additive exPlanations (SHAP) values to visualize and elucidate the significance of each factor involved. This study was reported using the TRIPOD checklist.

In the 7 ML models developed, both the Random Forest model and the XGBoost model exhibited exceptional performance. In addition, the study incorporated the SHAP framework to enhance the interpretability of these models. The importance of SHAP variables in the model from high to low were operation duration, age, serum albumin, body mass index, body temperature, and anesthesia grade.

The ML model developed in this study demonstrated strong performance in predicting perioperative PIs. Furthermore, the SHAP method offered valuable insights for interpreting the ML model.