Improving Sepsis Prediction in the ICU with Explainable Artificial Intelligence: The Promise of Bayesian Networks

Area of Science:

• Critical care medicine
• Artificial intelligence in healthcare
• Probabilistic modeling

Background:

• Sepsis is a major global cause of mortality with complex presentations.
• Early sepsis detection in ICUs is challenging due to data gaps and uncertainty.
• Current machine learning models often lack transparency, hindering clinical trust and adoption.

Purpose of the Study:

• To explore the advantages of Bayesian Networks (BNs) and Dynamic Bayesian Networks (DBNs) for sepsis prediction.
• To highlight the potential of interpretable AI in critical care settings.
• To bridge the gap between AI capabilities and bedside clinical practice.

Main Methods:

• Review of recent applications of probabilistic graphical models, specifically BNs and DBNs, for sepsis prediction.
• Analysis of AI models that explicitly represent clinical reasoning under uncertainty.
• Examination of real-time sepsis alert systems and treatment-effect modeling.

Main Results:

• DBNs achieved an AUROC of 0.94 for early sepsis detection.
• Causal probabilistic models showed an AUROC of 0.95 for hospital admissions.
• BNs and DBNs natively handle missing data and provide transparent decision paths.

Conclusions:

• BNs offer a transparent and interpretable alternative to opaque AI models for sepsis prediction.
• These models facilitate human-in-the-loop collaboration and integration into clinical workflows.
• Bayesian models provide the necessary performance and epistemic humility for data-driven critical care decisions.