M.M. Maurer, B. Pfitzner, R.P. van de Water, L. Faraj, C. Riepe, D. Zuluaga, F. Krenzien, N. Raschzok, R. Siegel, C. Schineis, B. Arnrich, K. Beyer, J. Pratschke, I.M. Sauer, and A. Winter evaluated federated learning (FL) as a privacy-preserving approach for AI-based prediction of 90-day mortality after colorectal surgery. Limited data sharing between hospitals often restricts surgical AI development, and FL allows multicenter model training without transferring raw patient data. The study also assessed the effect of differential privacy (DP) on model performance.
Data from 2,959 patients undergoing elective colorectal surgery at three tertiary centers (2015–2021) were analyzed. Neural networks were trained locally at each center, using centralized data aggregation and distributed federated learning. Additional privacy protection was implemented using central and local differential privacy.
Results showed that federated learning performed similarly to centralized modeling, achieving comparable predictive accuracy (AUROC ~0.78 vs. 0.81). However, adding differential privacy reduced performance, with central DP causing moderate declines and local DP nearly eliminating predictive accuracy. Across models, the most influential predictors were patient age, blood parameters, and the Charlson Comorbidity Index.
Overall, the study demonstrates that federated learning can enable effective multicenter surgical AI models while preserving data privacy, though strong privacy mechanisms like differential privacy may significantly compromise model performance.

"Privacy preserving federated learning for 90-day mortality prediction in colorectal surgery: a multicenter retrospective development and comparison study" is available in International Journal of Surgery 2025;111(12):9065-9074