Predicting Pressure Injury in Intensive Care Patients With Ensemble-Based Machine Learning Methods

This study was conducted to predict the risk of pressure injury in patients hospitalized in the intensive care unit using machine learning methods. Data of patients hospitalized in different intensive care units in the last year, such as surgical, general, internal medicine, coronary, cardiovascular surgery units, were collected. Preprocessing stages were applied to the data, and treatment prediction was performed with ensemble learning-based methods. Applied treatments were considered in 4 categories in total: (0) anti-microbial, (1) epithelialization, (2) protector, and (3) foam cover. The occurrence rates of the applied treatment classes were equalized with the oversampling method to avoid the imbalance problem in the data recorded. Afterward, training was performed by combining LightGBM, Random Forest, and AdaBoost models with an optimization approach. Random Forest showed the highest performance in classifying applied treatments with an overall accuracy of 0.76. Pressure ulcer stages were also analyzed in the study. This study demonstrates ensemble learning-based methods to successfully predict pressure injury treatment methods applied in intensive care units. The preprocessing steps and optimization methods involved in making clinical data suitable for machine learning models resulted in high accuracy of the training models. With this developed method, the staging of pressure injuries based on the individual characteristics of patients and personalized care recommendations for nursing care are presented. Machine learning can provide support for clinical decision-making in nursing care where more data is provided. These achievements emphasize using such machine learning-based models effectively in future nursing applications.