<p>Infections after surgery remain a leading cause of morbidity and mortality, yet reliable risk stratification at the end of surgery is limited. Intraoperative vital signs are continuously recorded in modern operating rooms but remain an underexploited source of real-time prognostic information. We developed and validated a machine-learning model integrating intraoperative vital-sign dynamics to predict postoperative infections immediately at the end of surgery. We extracted arterial blood pressure, heart rate, oxygen saturation, temperature, and end-tidal CO₂ time-series from a clinical data warehouse, transforming these signals into interpretable summary, trend, and distributional descriptors. Using routine data from 10,719 surgical procedures, models incorporating interpretable intraoperative time-series features achieved an AUROC of 0.88 (95% CI, 0.85–0.91) for infection prediction at the end of surgery, significantly outperforming models based on preoperative variables alone. Model predictions were calibrated across major procedure clusters and interpretable through SHAP-based feature attribution. Our results demonstrate that intraoperative time-series data encode signatures of cumulative surgical and physiological stress, revealing early and clinically actionable signals of postoperative infection risk and enable an explainable machine-learning framework for perioperative monitoring systems. This explainable approach moves risk assessment from delayed postoperative testing to immediate, digital decision support, ready for integration into perioperative monitoring systems.</p>