Early identification of acute respiratory distress syndrome (ARDS) in severe community-acquired pneumonia (SCAP) are crucial for reducing morbidity and mortality. This study focuses on developing an optimal prediction model based on clinical data and biomarkers to detect the risk of SCAP-associated ARDS in adult ICU patients.

This investigation, utilizing the MIMIC-IV database, enrolled 3807 patients with SCAP and randomly allocated them into a training set (n = 2664) for model development and a testing set (n = 1143) as an internal validation cohort to assess the model's predictive performance. The outcome was defined as the incidence of SCAP-associated ARDS. Baseline clinical and laboratory characteristics of the patients were obtained. Selection of characteristic variables was performed using LASSO regression, followed by the construction of ten ML models: LGBM, KNN, CatBoost, SVM, XGBoost, DesicionTree, NB, RF, KNNC, and MLP. The evaluation of model performance is conducted through various indicators such as ROC curves, calibration curve, DCA, accuracy, specificity, recall, presicion and F1 score.

Initially, 80 characteristic variables were collected, and then 67 of them were selected for the next analysis. After conducting a univariate regression analysis, 32 variables with P   <  0.1 were gathered. Through a multicollinearity analysis(VIF≥10), 3 variables were deleted. The remaining 29 variables were subjected to LASSO regression analysis, and 8 of the most significant characteristic variables (Charlson, Lactate, Stroke, Race, AG, ALB, Sepsis, ROX) were selected to build 10 prediction models using machine learning methods. In the training set, the AUROC of the prediction models were respectively LGBM(AUROC = 0.9770), KNN(AUROC = 0.9879), CatBoost(AUROC = 0.9827), SVM(AUROC = 0.8095), XGBoost(AUROC = 0.9999), DesicionTree(AUROC = 0.7611), NB(AUROC = 0.5759), RF(AUROC = 0.8068), KNNC(AUROC = 0.9879), and MLP(AUROC = 0.9661). While in the test set, the AUROC were LGBM(AUROC = 0.9355), KNN(AUROC = 0.6343), CatBoost(AUROC = 0.8745), SVM(AUROC = 0.7912), XGBoost(AUROC = 0.9466), DesicionTree(AUROC = 0.7791), NB(AUROC = 0.5765), RF(AUROC = 0.8074), KNNC(AUROC = 0.6343), and MLP(AUROC = 0.7386). A web calculator utilizing 8 crucial variables to predict SCAP-associated ARDS in adult ICU patients for ARDSML ^pred , is now available on our website ( ards/ ).

The prediction model constructed based on 8 characteristic variables selected by LASSO regression and using the XGBoost algorithm has excellent predictive performance in predicting the occurrence of SCAP-associated ARDS in adult ICU patients. This data-driven predictive model will help clinicians to make quick and accurate diagnosis.