Advancing Multicenter AKI Prediction: Benefits of Collaborative Models Over Localized Approaches

New evidence demonstrates that multicenter AKI prediction models significantly outperform locally trained single-center models for forecasting postoperative acute kidney injury (AKI) after cardiac surgery.

The study, published in npj Digital Medicine, analyzed 43,926 cardiac surgery cases from 23 hospitals in the Multicenter Perioperative Outcomes Group (MPOG). It compared traditional single-center models against pooled multicenter models and a novel federated model stacking approach across three AKI severity levels. Results consistently showed superior discrimination with multicenter strategies in both temporal and external validation.

Clear Performance Gains Across AKI Severities

Pooled multicenter AKI prediction models achieved AUCs of 0.856, 0.890, and 0.911 for AKI stages 1+, 2+, and 3+ respectively in temporal validation — markedly higher than the 0.770, 0.796, and 0.821 delivered by single-center models. These advantages held in external validation, where pooled models reached AUCs between 0.882 and 0.950.

Notably, no hospital achieved its best performance using a purely local model. Smaller-volume centers benefited the most, with several low-case hospitals unable to train viable local models yet reaching AUCs above 0.80 through multicenter AKI prediction approaches.

The analysis drew on high-quality perioperative data from 31 academic and community hospitals across 23 states. The cohort consisted of adult patients undergoing open cardiac surgery with cardiopulmonary bypass between 2014 and 2022. AKI was defined using KDIGO criteria based on postoperative serum creatinine changes. Gradient-boosted decision tree models served as the base learners, with the first postoperative creatinine value emerging as the strongest predictor across all approaches.

Learning curve analyses revealed that roughly half the performance benefit of a 23-hospital network could be achieved with just four hospitals. This finding has important implications for hospitals with limited case volumes that lack sufficient local data to develop robust predictive models independently.

Practical Path for Privacy-Preserving Collaboration

The study introduced a novel one-round federated model stacking method that trains local base models before combining their predictions through a meta-learner. This approach, which can optionally incorporate site-level metadata, offers a secure and efficient alternative to full data pooling while still delivering strong performance. It addresses growing concerns around data privacy and cybersecurity in healthcare AI development.

Strategic Implications for Health Systems and Value-Based Care

These results challenge the current emphasis on fully localized AI models in electronic health records. For health systems, payers, and technology developers, the evidence supports greater investment in multicenter AKI prediction infrastructure. Such collaboration can improve model discrimination, promote equity between large and small institutions, and enhance both clinical outcomes and cost-effectiveness in perioperative care.

The full study is available at npj Digital Medicine.