Federated Learning for Predicting Major Postoperative Complications.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Surgical Outcomes Research

Background:

• Predicting postoperative complications requires large, diverse datasets, often hindered by data privacy concerns.
• Artificial intelligence (AI) models offer potential for accurate risk prediction but face data access limitations.
• Federated learning presents a privacy-preserving method for collaborative model training across institutions.

Purpose of the Study:

• To develop a robust federated learning model for predicting postoperative complications.
• To ensure data privacy while utilizing clinical data from multiple institutions.
• To compare the performance of federated learning models against local and central models.

Main Methods:

• Retrospective cohort study of adult patients undergoing major surgery at two institutions (n=108,486).
• Development of federated learning models to predict nine major postoperative complications.
• Comparison of federated models with single-site local models and a pooled central model.

Main Results:

• Federated learning models achieved strong predictive performance (AUC 0.71-0.90) for various complications across both sites.
• Model performance was comparable to central models, with slight variations for specific outcomes like prolonged ICU stay.
• Federated models demonstrated performance similar to the best-performing local models.

Conclusions:

• Federated learning is a viable and effective tool for training robust postoperative outcome prediction models.
• This approach enables the use of large-scale, multi-institutional data while addressing privacy concerns.
• Federated learning facilitates collaborative AI development in healthcare for improved surgical risk assessment.