Abstract
Acute kidney injury (AKI) is a life-threatening problem for hospitalized patients, and early detection is crucial to reduce severe outcomes. Traditional predictive methods lack in monitoring complex physiological patterns and ensuring data privacy in decentralized healthcare settings. The study aims to develop and evaluate two distinctly complementary novel graph-based approaches, namely the centralized Graph Attention Network (GAT) and the decentralized model, Gossip Learning with Adaptive Aggregation GAT (GL-AA-GAT), to identify AKI onset between 6 and 12 h in advance, using physiological time-series data from the Kaggle Sepsis dataset. Multi-Head Attention is used for modeling feature interactions in centralized GAT, while GL-AA-GAT can further achieve this by decentralized training of five nodes using gossip exchange and adaptive aggregation for privacy and scalability. Through its novel graph structure, centralized GAT predicts the onset of AKI with an accuracy of 94.1%, sensitivity of 94%, AUC-ROC of 95%, and AUPRC of 91%. With decentralized privacy additions, GL-AA-GAT achieves an accuracy of 92.8%, a sensitivity of 93%, an AUC-ROC of 93.8%, and an AUPRC of 90%, with robustness. Sensitivity analyses revealed that the performance was stable with the prediction horizons and correlation thresholds, and external validation on non-sepsis cohorts of the ICU further indicated generalizability. Both models outperform existing models, which means high predictive reliability. The GL-AA-GAT's distributed approach gives privacy and flexibility, making it innovative for distributed clinical environments, with task scheduling enhancing training efficiency.