Construction And Validation Of Pressure Injury Risk Prediction Model For Intensive Care Unit Patients

BackgroundICU patients have a high incidence,high risk and low cure rate of pressure injuries.Carrying out structured and comprehensive pressure injury prevention and risk management is crucial to ensure patient safety and improve the quality of care.The advent of the big data era has pushed nursing management,nursing research and clinical care to change from traditional institutionalized management,data-based analysis and single application to precise management,intelligent analysis and diversified applications.Machine learning methods provide new ideas for pressure injury risk prediction.It has improved the efficiency of collecting and managing data related to pressure injury,realized the automatic analysis and processing of data,and promoted the management of pressure injury from the stage of“System management”to the stage of“Data management”and“Accurate management”.ObjectivesBased on the machine learning algorithm to construct a pressure injury risk prediction model in ICU patients,compare the predictive performance of each model,obtain the optimal model,and perform interpretability analysis.On the premise of improving the scientific accuracy of pressure injury risk prediction,it provides scientific basis and theoretical foundation for the subsequent research on big data related to pressure injury and the formulation of personalized risk management strategies,so as to achieve accurate prevention and management of pressure injury.MethodsObservational studies are used.The“Long hu hui”pressure injury mobile information platform was used as a research tool.To collect clinical data related to pressure injury from 17 tertiary hospitals in Gansu Province from April 01,2021 to October 31,2022.（1）SPSS 26.0 software was used to create a database for predicting the risk of pressure injury in ICU patients after pre-processing the data to describe the current occurrence of pressure injury in ICU patients.（2）The maximum correlation minimum redundancy（MRMR）,extreme random tree（ET）and recursive feature elimination（RFE）of the feature selection methods were used for feature selection of all variables,and the results of the three methods were combined to screen the final included variables.（3）Five algorithms of support vector machine,decision tree,random forest,plain Bayesian and K-nearest neighbour were used for the construction and validation of ICU patient pressure injury risk prediction models,and the accuracy,recall,precision,F1value,MCC value and AUC value of different models were compared to derive the optimal ICU patient pressure injury risk prediction model.（4）Interpretability analysis of the optimal ICU patient pressure injury risk prediction model using the SHAP tool.Results（1）Occurrence of pressure injury in ICU patientsThe overall incidence of pressure injury in 4704 ICU patients was 1.40%.A total of 93 sites occurred in 66 patients with pressure injuries.The most common site of occurrence was the sacrococcygeal region with 43（46.24%）,followed by the heel with10（10.75%）,and stage 2 pressure injuries accounted for the most with 60（62.50%）,followed by stage 1 with 22（22.92%）.（2）Construction and validation of a pressure injury risk prediction model for ICU patientsThe top 5 features of importance for MRMR screening were use of isoproterenol（15.167）,use of PICC（3.195）,cachexia（2.239）,stroke（2.159）,and single organ failure（2.078）.The top 5 features of importance for ET screening were education level（0.028）,nutritional problems（0.026）,abnormal skin colour（0.025）,p H（0.025）,and skin breakdown erythema（0.024）.The top 5 features of importance for RFE screening were BMI（0.037）,nutritional problems（0.030）,education level（0.029）,age（0.027）,and abnormal skin colour（0.026）.The variables identified by the three feature selection methods were combined,and two strategies incorporating subsets of the three groups of variables,intersection（F-I）and concatenation（F-U）,were used for the construction and validation of a predictive model of pressure injury risk in ICU patients.Among the risk prediction models constructed based on the two strategies,the random forest model had the best performance evaluation(AUC_F-I=0.874±0.106,Accuracy _F-I=0.990±0.007;AUC_F-U=0.816±0.101,Accuracy _F-U=0.990±0.007).Among them,the model constructed based on the F-I（length of ICU stay,gender,education level,smoking,weight loss in the last3 months,use of PICC,use of thoracic catheter,use of norepinephrine,use of reserpine,use of isoproterenol,use of pacifiers,mobility,nutritional problems,abnormal skin colour,unconsciousness,forced position,single organ failure,stroke,major surgery or trauma,arterial Pa O₂,heart rate）performed optimally and was more suitable for risk management of pressure injuries in ICU patients.（3）Interpretability analysis of an optimal ICU patient pressure injury risk prediction modelThe impact of decision weights and prediction directions of individual features within the optimal random forest model were visualized and interpreted through SHAP tool.The model was shown to run with abnormal skin colour,unconsciousness,length of ICU stays,and use of isoproterenol as important decision factors for the pressure injury risk prediction model.ConclusionsThis study screened variables associated with the occurrence of pressure injury in ICU patients based on three feature selection methods.Two strategies incorporating subsets of three groups of variables,intersection（F-I）and concatenation（F-U）,were used to construct and validate a prediction model for pressure injury risk in ICU patients.It was found that the random forest model constructed based on the F-I strategy performed optimally.Interpretable analysis of the model showed that abnormal skin colour,unconsciousness,length of ICU stays,and use of isoproterenol were important decision factors in the pressure injury risk prediction model.The risk prediction model constructed in this study can help clinical nurses better understand the factors that influence the occurrence of pressure injury in ICU patients,and thus provide a theoretical basis for nurses to accurately assess the risk of pressure injury and carry out targeted prevention and care.