SurgTrajecDiff: Deviation-aware perioperative risk assessment from patient-specific physiological trajectories in gynecologic oncology surgery

Perioperative deterioration during gynecologic oncology surgery is often reflected not by an isolated abnormal reading but by progressive deviation from the physiologic course expected for an individual patient. Existing perioperative risk models, however, are still largely based on static or short-horizon classification. We propose SurgTrajecDiff, a deviation-aware framework that learns patient-specific normative intraoperative trajectories from uncomplicated cases using a conditional diffusion model and quantifies real-time divergence via a neural stochastic differential equation. Severe postoperative complications were defined as Clavien–Dindo grade III or higher within 30 days. The framework was developed on 2845 patients and evaluated in an independent validation cohort of 1203 patients. SurgTrajecDiff achieved AUCs of 0.901 and 0.885 on the internal and external cohorts, compared with 0.871 and 0.856 for the strongest benchmark. For intraoperative hypotension warning, it provided a median lead time of 13.2 min, 84.7% sensitivity at 10 min, 86.3% specificity, and 0.5 false alerts/h. These results suggest that deviation from individualized expected physiologic trajectories can improve perioperative risk assessment and enable earlier warning, although prospective validation and deployment studies remain necessary.