Federated learning (FL)-driven real-time decision support for intraoperative cardiovascular surgery: A privacy-preserving AI framework

Intraoperative cardiovascular surgery demands fast and precise decision-making utilizing patient data in real time. Unfortunately, centralized artificial intelligence (AI) models have inherent drawbacks, including high latency, scalability such as low scalability, and privacy related to sharing raw patient data. This study outlines a federated learning (FL)-powered real-time decision support system (FDSS) capable of functioning over multiple distributed surgical nodes while ensuring the privacy of patient information. We developed a federated real-time surgical decision support (FRSDS) algorithm, which embodies the operational logic of the FDSS framework. The FDSS includes individual model training at the edge, differentially private updates, secure aggregation using Federated Averaging (FedAvg), and real-time inference at the surgical site. The framework was evaluated under multicenter cardiovascular datasets. We assess performance through accuracy, sensitivity, specificity, precision, F1-score, and area under the receiver operating characteristic curve (AUC-ROC), and system performance is assessed through latency and communication overhead. FDSS achieved 95.4% accuracy, 94.6% sensitivity, 95.0% specificity, 93.2% precision, and an AUC-ROC of 0.98. The averages exceeded those of centralized and baseline FL benchmark models. The average inference latency in real time was 120[Formula: see text]ms, which is appropriate for intraoperative use. The expected convergence was very strong; communication overhead was lessened with top-k gradient compression, and differential privacy was maintained with ([Formula: see text], [Formula: see text]) protection. The system proposed with the FDSS framework is an inoperative cardiovascular care option for supporting surgical performance that preserves patient privacy. And, because no raw patient data is transferred internally with the system, the FRSDS algorithm provides rapid inference and enables real-time, scalable decision-support policymaking capability even in intraoperative circumstances poised with patient risk.