The study "Overcoming the data barrier: transfer learning for 90-day mortality prediction in general surgery - a retrospective multicenter development and comparison study" was published in the January issue of the International Surgery Journal by A. Winter, B. Pfitzner, R.P. van de Water, L. Faraj, C. Riepe, W.H. Hahn, F. Krenzien, C. Schineis, T. Malinka, W. Schöning, C. Denecke, B. Arnrich, K. Beyer, J. Pratschke J, I.M. Sauer, and M.M. Maurer.

This multicenter study investigated whether transfer learning (TL) can improve AI-based prediction of 90-day postoperative mortality in general surgery, where limited datasets often hinder the development of robust models.

Data from 14,922 patients undergoing esophageal, liver, pancreatic, or colorectal surgery across three tertiary centers (2015–2023) were analyzed using 85 preoperative variables. Large source neural network models were first trained, then fine-tuned for specific surgical procedures using transfer learning. These models were compared with conventional machine learning (ML) approaches and standard clinical risk scores.
Results showed that ML models already outperformed traditional risk scores (e.g., ASA and Charlson Comorbidity Index). Transfer learning further improved performance, particularly in predicting mortality for esophageal (+38% AUPRC), liver (+14%), and pancreatic surgery (+8%). Across all models, patient age and the Charlson Comorbidity Index were the most influential predictors.
Overall, the study demonstrates that transfer learning can significantly enhance AI model performance in surgical settings with limited data, offering a promising strategy for improving preoperative risk stratification and decision-making in general surgery.