Real-time train rescheduling using exact mathematical optimization is challenged by the computational complexity of large-scale disruptions. This paper introduces a machine learning-based framework that reduces problem complexity by fixing a subset of decision variables prior to optimization. The approach fixes nominal route choices of trains by predicting infrastructure-related features, allowing for an effective reduction of the search space for exact solvers. Therefore, Random Forest and Logistic Regression classifiers are trained on generated disruption scenarios and corresponding high-quality rescheduling solutions. We evaluate the proposed framework on a real-world medium-sized station area in Germany. Results show that the method reduces computation time by well over 90 %, while maintaining high feasibility and competitive solution quality. This work demonstrates the potential of hybrid rescheduling methods for intelligent railway operations and offers a promising tool for real-time dispatching support.