Operating room (OR) scheduling in modern hospitals must reconcile elective case planning precision with emergency arrival uncertainty. Traditional two-phase approaches---offline elective and emergency case scheduling and emergency calling forth rescheduling in the reactive mode---are marred by computational delays and the wasteful use of resources. We introduce an integrated, real-time dynamic OR scheduling model grounded on deep reinforcement learning (DRL). By casting the OR scheduling problem as a Markov decision process, our method continuously adapts surgery assignments, start times, and overtime accruals to both elective and emergent requirements. An actor--critic structure trained with proximal policy optimization learns to progressively dispatch cases to minimize total overtime without separate rescheduling steps. On synthetic benchmarks for small, medium, and large instances with uniform and Poisson emergency-arrival distributions, the DRL outperformed a reactive mixed-integer programming (RMIP) baseline and an optimized heuristic. It reduced average overtime by over 50% while generating schedules in less than 4 s–achieving 100-1,000× speedup over RMIP. These results demonstrate that our learning-based method offers a scalable, reliable, and real-time solution to dynamic OR scheduling, paving the way for deployment in high-stakes clinical environments.