Abstract
Pressure injury (PI) is common in the ICU and not well captured by single-risk tools such as the Braden scale. We aimed to develop and internally validate a machine-learning model to predict new-onset PI using routinely collected ICU data. This retrospective single-centre cohort included adult ICU patients with length of stay ≥ 48 h (2018-2023). The primary outcome was new-onset PI during ICU stay. Candidate predictors were pre-specified: minimum albumin, maximum lactate, SOFA, APACHE II, first recorded Braden score, age, BMI, a nutrition score and treatment indicators. Missing values were imputed (median/mode). A gradient boosting model (GBM) was evaluated with stratified 3-fold cross-validation; a random forest (RF) served as a benchmark (stratified 70/30 train-test split). Discrimination (AUC) was primary; calibration, Brier score, decision-curve analysis (DCA) and feature importance were secondary. Logistic regression quantified independent associations. Among included ICU stays, 14.6% developed PI. On multivariable analysis, higher lactate, lower albumin, lower Braden scores, older age, CRRT, prone positioning, enteral nutrition and analgesic exposure were associated with increased PI risk, whereas sedatives showed an inverse association. The GBM achieved AUC≈0.69 with acceptable calibration and net clinical benefit across thresholds commonly used in preventive workflows (≈0.10-0.50). Single markers or simple combinations displayed only modest discrimination. A GBM built from routine ICU data provided moderate, well-calibrated discrimination for predicting new-onset PI and demonstrated decision-relevant net benefit. The model can complement Braden-based screening by refining risk stratification and prioritising intensified prevention for patients most likely to benefit. External validation and prospective evaluation are warranted.