Multi-organ spatiotemporal information aware model for sepsis mortality prediction

Area of Science:

• Computational biology and bioinformatics
• Medical informatics
• Artificial intelligence in healthcare

Background:

• Sepsis is a life-threatening syndrome of multi-organ dysfunction.
• Mortality in sepsis correlates with the number of affected organs.
• Existing prognosis models overlook crucial spatiotemporal interactions between organs, limiting their effectiveness, especially with sparse data.

Purpose of the Study:

• To develop a comprehensive model for predicting sepsis-induced in-hospital mortality.
• To capture complex spatiotemporal interactions among multiple organ systems.
• To improve sepsis prognosis, particularly with limited clinical data.

Main Methods:

• Developed a recurrent Graph Attention Network-multi Gated Recurrent Unit (rGAT-mGRU) model.
• Utilized parallel GRU sub-models for temporal organ variations and GAT for spatiotemporal connections.
• Employed an attention-injection mechanism for multi-organ data flow.
• Trained and tested on 10,181 sepsis cases from the MIMIC-III database.

Main Results:

• The rGAT-mGRU model achieved an AUROC of 0.8777 ± 0.0039 and AUPRC of 0.5818 ± 0.0071.
• Demonstrated superior performance compared to common baseline models.
• Effectively delineated organ contributions over time via attention weights.
• Maintained consistent performance with limited clinical data.

Conclusions:

• The rGAT-mGRU model shows potential for sepsis prognosis by analyzing dynamic spatiotemporal organ interplay.
• Offers valuable auxiliary decision-making support for clinicians in critical care settings.
• Enhances understanding of multi-organ system dynamics in sepsis.