Development of a neural network system for predicting in-hospital pressure ulcers in spinal trauma patients

This study aimed to developing and to deploying an optimal machine learning model to predict pressure ulcers (PUs) in hospitalized patients with spinal fractures, using data from the National Spinal Cord and Column Injury Registry of Iran (NSCIR-IR). Data were from 4326 patients with traumatic spinal fractures. The preprocessing phase was included handling missing values, feature engineering, normalization, and addressing class imbalance (with a 3.4 % PU incidence) using the Synthetic Minority Oversampling Technique (SMOTE). Feature selection was carried out with univariate filtering methods such as ANOVA and chi-square tests, along with the random forest feature importance algorithm. Six traditional machine learning (ML) algorithms and six ensemble models were trained and evaluated using metrics such as the area under the receiver operating characteristic curve (AUC), recall, and Brier score. The multilayer perceptron neural network (MLP) emerged as the top-performing model, offers advantages for clinical use due to a higher AUC of 0.888 (0.85–0.92), a balanced accuracy, a good recall, calibration, and an acceptable net benefit on the decision curve. Key predictors identified included the ASIA Impairment Scale, the Glasgow Coma Scale score, SCI type, SaO2, and the number of damaged vertebrae. Shapley Additive Explanations (SHAP) analysis further highlighted the directional influence of these factors on PU risk. The MLP model effectively predicts PU in patients with spinal fractures, outperforming other algorithms. Identified predictors align with clinical insights, are emphasizing the need for targeted preventive measures in hospitals. However, external validation with a larger multicenter cohort is recommended to confirm and to expand upon these findings. • This study compares the predictive performance of a diverse set of algorithms for pressure ulcers. • The multilayer perceptron (MLP) network outperformed other models, delivering stable and generalizable performance metrics. • AUC of 0.88 (0.85–0.92), balanced accuracy of 0.837 (CI: 0.78–0.88), recall of 0.862(CI: 0.76–0.94), Brier score of 0.13. • PU predictors are ASIA, spinal cord injury type, low consciousness, a high number of fractured vertebrae, and low SPO2.