AI-based disease severity grading predicts complications in laparoscopic appendectomy

Appendicitis severity underpins contemporary management guidelines, where laparoscopic appendectomy remains gold standard. Preoperative measures poorly predict actual disease severity or complication risk, while operative grading systems such as the American Association for the Surgery of Trauma (AAST) remains largely confined to research settings. Artificial intelligence (AI) may provide practical solutions. We evaluated a previously validated AI-derived surgical video assessment of disease severity for predicting perioperative complications.

This retrospective study included consecutive surgical videos (6/2022-1/2024) routinely analyzed by the AI platform. AI-derived severity scores were stratified into Low (uncomplicated) and High (complicated) groups. Multivariable analysis identified independent predictors of complications. Operative-AAST served for benchmarking. Model discrimination (AUC), post hoc recalibration plot, and decision curve analysis (DCA) were evaluated.

Of 632 cases, 74.5% were low severity and 25.5% high. The High group had higher complication rates (26.7% vs. 10%; p<0.001), including intraoperative (11.8 vs. 3.2%; p < 0.001) and postoperative complications (15.5 vs. 7.5%; p = 0.005). AI-derived severity independently predicted complications (OR 2.76, 95% CI 1.62-4.73; p < 0.001), even after adjustment for operative-AAST (OR 1.90, 95% CI 1.07-3.36; p = 0.028). Discrimination was modest (AUC = 0.63), similar to operative-AAST (AUC = 0.68). Calibration plot showed incremental increase in complication rates across probabilities, with acceptable agreements at extremes and improved alignment at intermediate-risk. DCA showed the model had highest net benefit at intermediate-risk, comparable or higher than operative-AAST.

Automated AI-based surgical video assessment shows promise as a complementary tool for risk-prediction of laparoscopic appendectomy. It offers scalable risk stratification that may be implemented in routine clinical practice. Nevertheless, further study and model refinement are warranted.