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

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Graph Neural Networks for Predictive Analytics

Background:

• Acute kidney injury (AKI) poses a significant threat to hospitalized patients, necessitating early detection to mitigate severe outcomes.
• Traditional predictive models struggle with complex physiological data and data privacy in decentralized systems.
• Graph-based approaches offer a promising avenue for analyzing complex time-series physiological data.

Purpose of the Study:

• To develop and evaluate two novel graph-based models, a centralized Graph Attention Network (GAT) and a decentralized Gossip Learning with Adaptive Aggregation GAT (GL-AA-GAT), for early AKI detection.
• To predict AKI onset 6-12 hours in advance using physiological time-series data.
• To assess the performance and generalizability of these models, particularly the privacy-preserving decentralized approach.

Main Methods:

• Utilized the Kaggle Sepsis dataset, a physiological time-series dataset.
• Developed a centralized Graph Attention Network (GAT) employing Multi-Head Attention for feature interaction modeling.
• Created a decentralized GL-AA-GAT model using gossip exchange and adaptive aggregation for privacy-preserving, scalable training across multiple nodes.

Main Results:

• Centralized GAT achieved 94.1% accuracy, 94% sensitivity, 95% AUC-ROC, and 91% AUPRC.
• Decentralized GL-AA-GAT demonstrated strong performance with 92.8% accuracy, 93% sensitivity, 93.8% AUC-ROC, and 90% AUPRC, while ensuring privacy.
• Both models outperformed existing methods, indicating high predictive reliability and robustness across prediction horizons and external validation cohorts.

Conclusions:

• The developed graph-based models, GAT and GL-AA-GAT, offer highly reliable and accurate early detection of acute kidney injury.
• The decentralized GL-AA-GAT model presents an innovative, privacy-preserving, and flexible solution for distributed clinical environments.
• These advanced AI approaches hold significant potential for improving patient care and outcomes in critical care settings.