SBC-SHAP: Increasing the Accessibility and Interpretability of Machine Learning Algorithms for Sepsis Prediction.

Area of Science:

• Medical Informatics
• Computational Biology
• Machine Learning

Background:

• Sepsis is a leading global cause of mortality, necessitating early detection for effective treatment.
• Complete blood count (CBC) parameters offer potential as early sepsis indicators.
• Existing machine learning (ML) approaches for sepsis prediction lack interpretability and clinical accessibility.

Purpose of the Study:

• To develop an interpretable and accessible ML model for early sepsis prediction using CBC data.
• To improve ML model performance by integrating time-series information and reference ratios.
• To create a user-friendly web application for clinical deployment.

Main Methods:

• A graph-based ML framework was employed to process time-series CBC data.
• The impact of incorporating ratios relative to healthy reference values was evaluated.
• A web application, SBC-SHAP, was developed for visualizing sepsis risk and model interpretations.

Main Results:

• The novel approach improved sensitivity at 80% specificity for sepsis prediction, increasing it from 78.2% to 82.9% (internal) and 65.4% to 73.4% (external).
• The method's performance gains were consistent regardless of the specific ratio calculation.
• The SBC-SHAP web application provides interpretable insights into ML-based sepsis risk assessment.

Conclusions:

• The developed tool enhances the interpretability and accessibility of ML models for sepsis prediction.
• SBC-SHAP facilitates clinical adoption of advanced ML techniques for sepsis management.
• The open-source nature of SBC-SHAP promotes further research and development.