The study analyzed 482 patients with non-pulmonary sepsis for early prediction of ARDS using machine learning algorithms based on inflammatory indicators and blood gas parameters. Using the Recursive Feature Elimination (RFE) method, 11 key variables were selected and used to build nine machine learning models. The LightGBM model achieved high accuracy with an AUC of 0.954 in the training group and 0.923 in the test group, while its calibration curve indicated the reliability of the predictions. Decision curve analysis (DCA) confirmed the clinical value of the LightGBM model with the highest net gain in the 0–0.4 threshold range. The most important predictors of ARDS were SOFA score, PaO2/FiO2 ratio, lactate level, creatinine and SAPS II score according to SHAP analysis. Thus, the model successfully predicts the risk of ARDS in patients with non-pulmonary sepsis, which may help in early intervention and improve the prognosis.