Development and External Validation of a Prediction Model for Postoperative Pulmonary Complications in Mechanically Ventilated Surgical Patients Using MIMIC-IV and eICU

Background Postoperative pulmonary complications (PPC) substantially affect surgical outcomes. We aimed to develop and externally validate a risk prediction model for postoperative pulmonary complications in mechanically ventilated surgical patients using public critical care databases. Methods The development cohort was derived from the Medical Information Mart for Intensive Care (MIMIC)-IV, including adults with documented surgery who received invasive mechanical ventilation within 24 h of intensive care unit admission; external validation used the eICU Collaborative Research Database. The primary outcome was a narrow composite of atelectasis, pneumonia, or aspiration pneumonitis. Multiple imputation and least absolute shrinkage and selection operator (LASSO) regression were used; internal validation used 10-fold cross-validation (50 repeats), and external validation assessed discrimination and calibration. Results The development cohort comprised 18379 patients with 4650 PPC events (25.3%); the external cohort comprised 2778 patients with 549 events (19.8%). The final model included 16 predictors, with an area under the receiver operating characteristic curve (AUC) of 0.731 in development and 0.729 on internal cross-validation. External validation yielded an AUC of 0.680 (95% confidence interval 0.655–0.706) and a Brier score of 0.147. Adding vital signs and ventilator parameters to a base clinical model improved the AUC from 0.667 to 0.731. Conclusion The model showed moderate discrimination and was externally validated in an independent database (AUC 0.680). Its structure was transportable, but absolute risk required recalibration in the external cohort; the model is therefore better suited to risk stratification than to individual absolute-risk prediction.