Integrating Process Mining and Machine Learning for Surgical Workflow Optimization: A Real-World Analysis Using the MOVER EHR Dataset

The digitization of healthcare has enabled the application of advanced analytics, such as process mining and machine learning, to electronic health records (EHRs). This study aims to identify workflow inefficiencies, temporal bottlenecks, and risk factors for delayed recovery in surgical pathways using the open-access MOVER dataset. A multi-stage framework was implemented, including heuristic control-flow discovery, Petri net-based conformance checking, temporal performance analysis, unsupervised clustering, and Random Forest-based classification. All analyses were simulated on pre-discharge (“preliminary”) patient records to enhance real-time applicability. Control-flow models revealed deviations from expected pathways and issues with data quality. Conformance checking yielded perfect fitness (1.0) and moderate precision (0.46), indicating that the model generalizes despite clinical variability. Stratified performance analysis exposed duration differences across ASA scores and age groups. Clustering revealed latent patient subgroups with distinct perioperative timelines. The predictive model achieved 90.33% accuracy, though recall for delayed recovery cases was limited (24.23%), reflecting class imbalance challenges. Key features included procedural delays, ICU status, and ASA classification. This study highlights the translational potential of integrating process mining and predictive modeling to optimize perioperative workflows, stratify recovery risk, and plan resources.