Novel graph-based centralized and decentralized approaches for early AKI prediction.

Summary

The authors propose centralized and decentralized graph attention models that predict AKI 6–12 hours before onset using ICU time-series from a sepsis dataset, achieving AUC-ROC up to 0.95 and AUPRC 0.91. A decentralized gossip learning variant preserves privacy while maintaining high performance and robustness, with external validation and sensitivity analyses supporting generalizability.

Key Findings

• Centralized GAT predicted AKI 6–12 hours ahead with accuracy 94.1%, sensitivity 94%, AUC-ROC 95%, and AUPRC 91%.
• Decentralized GL-AA-GAT achieved accuracy 92.8%, sensitivity 93%, AUC-ROC 93.8%, and AUPRC 90% with privacy-preserving training across five nodes.
• Performance was robust across prediction horizons and correlation thresholds, and external validation on non-sepsis ICU cohorts supported generalizability.
• Both models outperformed existing baselines.

Clinical Implications

If prospectively validated, such models could enable earlier nephrology consults, targeted hemodynamic and nephrotoxin stewardship, and resource allocation without centralizing data.

Limitations

• Retrospective modeling on public datasets; potential dataset shift and selection biases
• Lack of prospective, real-time clinical deployment and impact assessment

Future Directions

Prospective, multi-center silent trials and randomized implementation studies to assess clinical impact on AKI incidence, sepsis outcomes, and nephrotoxin stewardship under privacy-preserving federated settings.