An Explainable Artificial Intelligence Predictor for Early Detection of Sepsis

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Critical Care Medicine

Background:

• Early sepsis detection is vital to prevent irreversible organ damage and reduce mortality.
• Delayed treatment significantly increases sepsis-related mortality rates.
• Electronic Health Records (EHRs) contain rich data for developing predictive models.

Purpose of the Study:

• To develop an explainable artificial intelligence (AI) model for early sepsis prediction.
• To analyze EHR data for real-time sepsis risk assessment.
• To enhance the interpretability of AI models in clinical practice.

Main Methods:

• A retrospective observational study utilizing EHR data from the PhysioNet/Computing in Cardiology Challenge 2019.
• Development of an explainable AI sepsis predictor model trained on hourly extracted features (168 features).
• Model validation on shared public ICU data and full hidden test sets from multiple hospital systems.

Main Results:

• The explainable AI model demonstrated real-time sepsis prediction capabilities.
• Feature impact analysis provided interpretability for sepsis risk factors.
• The model achieved a clinical utility score of 0.364 on hidden test sets, with scores of 0.430 and 0.422 on two separate test sets.

Conclusions:

• The explainable AI sepsis predictor model offers superior performance for real-time sepsis risk prediction.
• The model provides interpretable insights into sepsis risk, aiding clinical understanding.
• This approach supports timely interventions and improved patient management in intensive care units (ICUs).