Abstract
We study a surgery scheduling problem under limited downstream ICU capacity, where the scheduling decisions are made before postoperative ICU demand is realized, while ICU admissions adapt to patient-specific feature information revealed after surgery. This temporal mismatch creates a misalignment between upstream surgical decisions and downstream ICU capacity, increasing congestion risk and operational inefficiency. To address this challenge, we develop a feature-driven robust optimization framework that integrates predictive modeling with operational decision-making. Our proposed model not only improves surgery-duration prediction but also preserves tractability and interpretability for optimization. Furthermore, to account for downstream congestion risk, we construct a feature-driven representation of ICU demand and formulate a distributionally robust capacity constraint. We show that this constraint can be equivalently transformed into a tractable Conditional Value-at-Risk measure, providing a risk-averse safeguard that effectively limits ICU congestion arising from adverse realizations of length-of-stay uncertainty and patient mix. Numerical studies using a comprehensive data set from a cardiology department in China demonstrate that our approach consistently outperforms traditional predict-then-optimize frameworks. In particular, our framework substantially reduces ICU bed shortages during peak-demand periods while improving operational robustness and enhancing the coordination between surgery scheduling and ICU capacity management.