Federated learning for AKI prediction in COVID-19 patients

Background: Predictive models are trained on single-center data and are nongeneralizable, and multi-center data pooling raises privacy concerns. Federated learning (FL) trains models by updating parameters from a central aggregator without sharing raw data. We used FL to predict acute kidney injury (AKI) in COVID-19 patients within 3 (AKI3) and 7 (AKI7) days of admission as a use case. Methods: We selected 4035 COVID-19 patients admitted to 5 hospitals in New York City, after excluding patients with kidney failure, to train logistic regression and logistic regression with L1 regularization (LASSO) models through 3 approaches: local data, pooled data from all sites, and a FL method. Results: Federated models outperformed local models as measured by area under the receiver operating characteristic curve (Figure 1, Table 1). SHAP plots indicate differences in feature importance between LASSO models in AKI3 prediction (Figure 2). Conclusions: FL has utility for developing accurate predictive models without compromising patient data. Average model performance across hospitals by area under the receiver-operating characteristic curve. SHAP plots of LASSO local and federated models in predicting AKI within 3 days of admission at Mount Sinai Hospital.