Exploring the limits of localization: federated model stacking improves hospital-level prediction in a national research network

Challenges with model generalizability and data privacy have led to a shift in health artificial intelligence (AI) models being trained locally within individual health systems rather than relying on multicenter data. Localization carries the promise of capturing local practice patterns and patient demographics, presumably resulting in better models. Our study empirically tests this hypothesis in a national research network by comparing locally trained models predicting acute kidney injury (AKI) after cardiac surgery with two multicenter modeling approaches, pooling and a novel federated model stacking method. Trained on 43,926 cases across 23 hospitals, the study finds that multicenter models outperform single-center approaches, with higher area under the receiver operating characteristic curves (AUCs) for all AKI severity levels in both temporal and external validation sets. Hospitals with smaller case volumes benefit the most from multicenter approaches, showing the greatest AUC increase over locally trained models.