Machine learning predicts acute respiratory failure in pancreatitis patients: A retrospective study

• Published in: International journal of medical informatics (Shannon, Ireland) Vol. 192; p. 105629
• Main Authors: Zhou, Liu-xin, Zhou, Qin, Gao, Tian-ming, Xiang, Xiao-xing, Zhou, Yong, Jin, Sheng-jie, Qian, Jian-jun, Zhou, Bao-huan, Bai, Dou-sheng, Jiang, Guo-qing
• Format: Journal Article
• Language: English
• Published: Ireland Elsevier B.V 01.12.2024
• Subjects: Acute pancreatitis; Acute respiratory failure; Machine learning; MIMIC- IV database; Prediction model; Acute respiratory failure; Acute pancreatitis; MIMIC- IV database; Machine learning; Prediction model
• ISSN: 1386-5056; 1872-8243; 1872-8243
• DOI: 10.1016/j.ijmedinf.2024.105629

•Ca, ALB, GLR, WBC, AG and BUN were predictors of ARF in AP patients.•The emergence of ARF in patients with AP is closely related to systemic inflammatory status, hypovolemia and internal environmental disturbances.•The XGBoost model has good differentiation ability and provides an important reference for assessing the risk level of ARF in patients.•ARF is the most frequent complication of AP, and ARF is a major contributor to the poor prognosis of AP. The purpose of the research is to design an algorithm to predict the occurrence of acute respiratory failure (ARF) in patients with acute pancreatitis (AP). We collected data on patients with AP in the Medical Information Mart for Intensive Care IV database. The enrolled observations were randomly divided into a 70 % training cohort and a 30 % validation cohort, and the observations in the training cohort were divided into ARF and non-ARF groups. Feature engineering was conducted using random forest (RF) and least absolute shrinkage and selection operator (LASSO) methods in the training cohort. The model building included logistic regression (LR), decision tree (DT), k-nearest neighbours (KNN), naive bayes (NB) and extreme gradient boosting (XGBoost). Parameters for model evaluation include receiver operating characteristic (ROC) curve, precision-recall curve (PRC), calibration curves, positive predictive value (PPV), negative predictive value (NPV), true positive rate (TPR), true negative rate (TNR), accuracy (ACC) and F1 score. Among 4527 patients, 445 patients (9.8 %) experienced ARF. Ca, ALB, GLR, WBC, AG and BUN have been included in the prediction model as features for predicting ARF. The AUC of XGBoost were 0.86 (95 %CI 0.84–0.88) and 0.87 (95 %CI 0.84–0.90) in the training and validation cohorts. In the training cohort, XGBoost demonstrates a true positive rate (TPR) of 0.662, a true negative rate (TNR) of 0.884, a positive predictive value (PPV) of 0.380, a negative predictive value (NPV) of 0.960, an accuracy (ACC) of 0.862, and an F1 score of 0.483. In the validation cohort, XGBoost shows a TPR of 0.620, a TNR of 0.895, a PPV of 0.399, an NPV of 0.955, an ACC of 0.867, and an F1 score of 0.486. The XGBOOST model demonstrates good discriminatory ability, which enables clinicians to ascertain the probability of developing ARF in AP patients.