Development and validation of a machine learning algorithm–based risk prediction model of pressure injury in the intensive care unit

Abstract

The study aimed to establish a machine learning-based scoring nomogram for early recognition of likely pressure injuries in an intensive care unit (ICU) using large-scale clinical data. A retrospective cohort study design was employed to develop and validate a top-performing clinical feature panel accessibly in the electronic medical records (EMRs), which was in the mode of a quantifiable nomogram. Clinical factors regarding demographics, admission cause, clinical laboratory index, medical history and nursing scales were extracted as risk candidates. The performance improvement was based on the application of the machine learning technique, comprising logistic regression, decision tree and random forest algorithm with five-fold cross-validation (CV) technique. The comprehensive assessment of sensitivity, specificity and the area under the receiver operating characteristic curve (AUROC) was considered in the evaluation of predictive performance. The receiver operating characteristic curves revealed the top performance for the logistic regression model in respect to machine learning improvement, achieving the highest sensitivity and AUC among three types of classifiers. Compared against the 23-point Braden scale routinely recorded online, an incorporated nomogram of logistic regression model and Braden scale achieved the best performance with an AUC of 0.87 ± 0.07 and 0.84 ± 0.05 in training and test cohort, respectively.Our findings suggest that the machine learning technique potentiated the limited predictive validity of routinely recorded clinical data on pressure injury development during ICU hospitalisation. Easily accessible electronic records held the potentials to substitute the traditional Braden score in the prediction of pressure injury in intensive care unit. Preoperative prediction of pressure injury facilitates the exemption from the severe consequences.

An Integrated System of Braden Scale and Random Forest Using Real-Time Diagnoses to Predict When Hospital-Acquired Pressure Injuries (Bedsores) Occur

“…Sixty variables that include the Braden Risk Assessment subscales were used as inputs for a model to predict HAPI timing. The HAPI timing was identified as the number of days to develop HAPIs [ 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 46 ]. These variables are shown in Table 1 and selected through literature survey and clinician’s feedback [ 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 <...…”

“…The model training and testing was repeated for 50 iterations; at every iteration, the training/testing data split was performed randomly. The RF model performance was compared with seven commonly applied algorithms in HAPI research that include: SVM, MLP, KNN, LDA, DT, LR, and Adaptive Boosting (AdaBoost) [ 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 ].…”

“…Machine Learning (ML) has been utilized to predict HAPI patients before occurrence using patients’ Electronic Health Records (EHR) as an adjunct to clinical assessment, which can further narrow down which patients are at risk [ 6 , 7 , 8 , 9 ]. In the last 15 years, 30 studies have answered who will develop HAPIs by utilizing classic machine and deep learning algorithms [ 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 ]. Two studies adopted Grid Search (GS) to optimize the hyperparameters of ML [ 10 , 22 ].…”

“…Two studies adopted Grid Search (GS) to optimize the hyperparameters of ML [ 10 , 22 ]. In most studies, random oversampling, undersampling, and Synthetic Minority Oversampling Techniques (SMOTE) were adopted as oversampling techniques [ 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 38 ]. Nevertheless, one study adopted cost-sensitive learning [ 22 ] to deal with the unbalanced dataset.…”

“…Nevertheless, one study adopted cost-sensitive learning [ 22 ] to deal with the unbalanced dataset. The most used algorithms were Random Forest (RF) [ 10 , 12 , 13 , 14 , 15 , 16 , 19 , 21 , 22 , 25 , 30 , 31 , 32 , 36 , 38 , 39 ], Logistic Regression (LR) [ 10 , 11 , 14 , 16 , 18 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 29 , 30 , 31 , 32 , 34 , 36 ], Decision Tree (DT) [ 15 , 16 , 21 , 23 , 25 , 26 , 27 , 28 , 31 , 32 , 36 ], Support Vector Machine (SVM) [ 10 , 12 , 21 , 22 , 25 , 26 , 27 , 28 , 30 , ...…”