基于机器学习构建急性胰腺炎患者并发急性呼吸窘迫综合征风险预测模型OA

Objective To develop a machine learning based risk prediction model for acute respiratory distress syndrome(ARDS)in patients with acute pancreatitis(AP)using routinely available clinical indicators,thereby providing decision support for clinicians.Methods A retrospective cohort study was performed based on the Medical Information Mart for Intensive Care Ⅳ(MIMIC Ⅳ)3.1 database.Patients with AP were identified and stratified according to the occurrence of ARDS.Logistic regression analysis was used to screen candidate predictors.Based on the selected feature subset,8 machine learning models were constructed,including Logistic regression,gradient boosting machine(GBM),adaptive boosting(AdaBoost),random forest,K-nearest neighbors(KNN),neural network,extreme gradient boosting(XGBoost),and support vector machine(SVM).Hyperparameters were optimized according to the characteristics of each algorithm.Model performance was comprehensively evaluated to identify the optimal model,which was subsequently interpreted using Shapley Additive Explanations(SHAP)analysis.Results A total of 1 553 patients with AP were included,among whom 616(39.67%)developed ARDS after intensive care unit(ICU)admission.Logistic regression with backward elimination identified 8 key predictors associated with ARDS in patients with AP:acute physiology and chronic health evaluation Ⅱ(APACHE Ⅱ)score,sequential organ failure assessment(SOFA)score,respiratory rate(RR),serum creatinine(SCr),blood glucose,pulse oxygen saturation(SpO2),and albumin(Alb)[odds ratios(OR)with 95%confidence intervals(95%CI)were 1.061(1.030-1.093),1.185(1.130-1.242),1.047(1.007-1.090),0.962(0.949-0.974),0.587(0.459-0.752),1.559(1.339-1.814),and 1.002(1.000-1.004)],P<0.001,<0.001,0.021,<0.001,<0.001,<0.001,0.049 respectively.Among the 8 machine learning models,the best performance was achieved by the XGBoost and AdaBoost models,with areas under the receiver operator characteristic curve(AUC)and 95%CI of 0.855(0.827-0.883)and 0.844(0.815-0.872),respectively.In the test set,these models maintained stable performance,with AUC values of 0.838(95%CI:0.795-0.881)for XGBoost and 0.832(95%CI:0.789-0.876)for AdaBoost,and both outperforming the other algorithms in both predictive accuracy and stability.SHAP analysis indicated that the most influential predictors included APACHEⅡ score,SOFA score,maximum serum creatinine,minimum albumin,and maximum blood glucose.Conclusions The XGBoost and AdaBoost models constructed using routinely available clinical indicators and scoring systems demonstrated good predictive performance and stability.These models may effectively identify AP patients at high risk of developing ARDS,facilitating early detection,risk stratification,and personalized clinical intervention.