Federated Learning for Predicting Major Postoperative Complications
Objective:. To develop a robust model to accurately predict the risk of postoperative complications using clinical data from multiple institutions while ensuring data privacy. Background:. Building accurate, artificial intelligence models to predict postoperative complications relies on accessibilit...
Saved in:
| Main Authors: | , , , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wolters Kluwer Health
2025-06-01
|
| Series: | Annals of Surgery Open |
| Online Access: | http://journals.lww.com/10.1097/AS9.0000000000000573 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Objective:. To develop a robust model to accurately predict the risk of postoperative complications using clinical data from multiple institutions while ensuring data privacy.
Background:. Building accurate, artificial intelligence models to predict postoperative complications relies on accessibility of large-scale and diverse datasets, often restricted by privacy concerns.
Methods:. This retrospective cohort study includes adult patients admitted to University of Florida Health (UFH) hospitals in Gainesville (GNV) (n = 79,850) and Jacksonville (JAX) (n = 28,636) for all inpatient major surgical procedures. We developed federated learning models to predict 9 major postoperative complications and compared them with both local models trained on a single site and central models trained on a pooled dataset from 2 hospitals.
Results:. Our best-federated learning models using preoperative features achieved the area under the receiver operating characteristics curve values with 95% confidence interval (CI) ranging from 0.80 (95% CI, 0.79–0.80) for wound complications to 0.90 (95% CI, 0.90–0.91) for prolonged intensive care unit (ICU) stay at UFH GNV. At UFH JAX, these values ranged from 0.71 (95% CI, 0.70–0.72) for wound complications to 0.90 (95% CI, 0.88–0.92) for in-hospital mortality. Federated learning models achieved comparable discrimination to central models for all outcomes, except prolonged ICU stay, where the performance of the federated learning model was slightly better at UFH GNV and slightly worse at UFH JAX. Our federated learning models obtained comparable performance to the best local models.
Conclusions:. We show federated learning to be a useful tool to train robust postoperative outcome prediction models from large-scale data across 2 hospitals. |
|---|---|
| ISSN: | 2691-3593 |