Federated learning to optimize coronary revascularization decisions using multi-center emergency room data: a privacy-preserving approach
European Heart Journal - Digital Health
Abstract
Coronary revascularization significantly improves outcomes in acute coronary syndrome (ACS) but remains challenging to predict accurately in emergency settings. Federated learning (FL) presents a promising approach to enhance predictive accuracy by securely integrating data from multiple hospitals without compromising patient privacy.
This study aimed to develop and validate an FL model that accurately predicts the necessity for coronary revascularization using initial emergency room (ER) data from multiple hospitals while ensuring data confidentiality.
We conducted a retrospective analysis using anonymized patient data (n=75,497) from three tertiary hospitals. The dataset was split into training (January 2017–July 2020) and test sets (August 2020–July 2021). Predictive models were developed using local learning (LL), centralized learning (CL), and federated learning (FL) enhanced by local differential privacy (LDP) (Fig 1). Model performance was assessed using AUROC, AUPRC, precision, recall, and accuracy. Clinical utility was evaluated by comparing model predictions with physician decisions using Cohen's kappa.
The FL model achieved high predictive accuracy, comparable to the CL model, with an average AUROC of 0.95 and AUPRC of 0.64 across hospitals. Incorporation of local differential privacy did not significantly impair performance (AUROC 0.95, AUPRC 0.64). Substantial agreement was observed between FL model predictions and physician decisions (kappa=0.654), especially in patients with elevated troponin levels (kappa=0.675). Furthermore, the FL model facilitated a strategic balance between reducing unnecessary coronary angiography procedures and maintaining high sensitivity for required interventions.
Our federated learning approach effectively predicts coronary revascularization requirements in emergency settings, offering enhanced decision support without compromising patient data privacy. This model demonstrates significant potential for optimizing clinical resource use and improving patient outcomes in cardiovascular emergencies.
