Abstract Postoperative pneumonia (POP) is a prevalent, severe complication in elderly noncardiac surgical patients, linked to extended hospital stays, increased healthcare costs, and higher mortality. Existing predictive models are often limited by single-center data, small cohorts, or restricted variables, highlighting the need for a comprehensive tool integrating multi-phase perioperative factors. This retrospective study analyzed 44,740 patients aged ≥ 65 years who underwent noncardiac surgery (November 2014–April 2022) at Henan Provincial People’s Hospital, with 3187 (7.1%) developing POP. Patients were stratified into development (n = 31,320) and validation (n = 13,420) cohorts via 70:30 random split. Key predictors were identified using LASSO logistic regression (from 44 candidates), followed by multivariate logistic regression with forward stepwise selection. Model performance was evaluated via AUC (discrimination), calibration (Hosmer–Lemeshow test, Brier score), clinical utility (decision curve analysis DCA), and interpretability (SHAP analysis). The final model included 9 predictors: anesthesia duration, anesthesia type, smoking status, pulmonary disease history, intraoperative colloid volume, preoperative anticoagulant/antihypertensive/steroid use, and intraoperative sufentanil dose. It demonstrated strong discrimination (validation AUC = 0.804, 95% CI 0.790–0.818) and good calibration (development: Hosmer–Lemeshow χ 2 = 5.45, P = 0.79; validation: χ 2 = 7.81, P = 0.55; Brier score = 0.058 for both). A derived nomogram (optimal cutoff = 190) showed high sensitivity (76.3%) and specificity (69.6%). DCA confirmed net benefit across 0–89% (development) and 0–88% (validation) thresholds. SHAP analysis identified prolonged anesthesia and pulmonary disease history as top predictors. This multifactorial model reliably predicts postoperative pneumonia in elderly noncardiac surgical patients using routinely collected perioperative data, with good discrimination and calibration. By integrating a wider range of variables than prior models, it enhances predictive accuracy and clinical applicability. External validation in multicenter prospective cohorts is needed to confirm its generalizability and support clinical integration.
Cong et al. (2025) studied this question.