Dynamic pressure ulcer risk predictions for hospitalized patients: development and validation of a machine learning model with an expert group of nurses

BackgroundPressure ulcers (PU) are a source of harm and discomfort in hospitalized patients. To identify patients at risk for developing PU nurses use validated risk assessment instruments which have insufficient predictive power, also leading to increased nursing workload. Machine learning on electronic health records (EHR) has the potential to provide better risk predictions while also reducing nursing workload. ObjectiveTo develop and validate a dynamic prediction model for daily PU risk predictions together with an expert group of nurses, while considering factors important to successful clinical implementation like explainability and robustness during the modelling process. MethodsAll admissions to general wards in a single-center tertiary university hospital in the Netherlands were included. A retrospective dataset with candidate predictors was collected from the EHR and split into a training set (December 2021 - November 2022, N=19931) and a validation set (January 2023 - June 2023, N=11387). Separate models were developed for the first 72 hours of admission, and >72 hours. The PU outcome was identified from both structured and free text registration. Predictor definition, selection and modelling choices were discussed with an expert group of nurses. Several candidate machine learning models were compared using cross-validation on the training set. The same final model and predictor set were selected for both [≤]72h and >72h, and predictive performance on the validation set was compared to the Waterlow score. ResultsThe difference in performance across candidate models was small on the training set. Logistic regression with an L2 penalty and a spline transformation applied to a limited set of predictors was chosen as the final model, and the candidate predictor set was reduced to a final predictor set. The AUROC for the final model was 0.790 ([≤]72h) and 0.795 (>72h) on the validation dataset. The model clearly outperformed the Waterlow score (0.816 vs. 0.702 ([≤]72h) and 0.800 vs. 0.677 (>72h)) for the subset of admissions with a registered Waterlow score. Expected remaining length of stay, age and ADL (activities of daily living) score were the predictors with the strongest contribution. ConclusionsWe developed and validated a dynamic prediction model for daily PU risk predictions, which outperformed the traditional Waterlow risk assessment score. During the development process special care was given to considerations for implementation and user acceptance. The model was named DRAAI (Decubitus Risk Alert based on AI), which means turn in Dutch and is a common PU preventive measure.