Prediction of Sepsis in the Intensive Care Unit With Minimal Electronic Health Record Data: A Machine Learning...

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Sepsis is a leading cause of hospital mortality, with current prediction methods lacking reliability and speed.
• Early and accurate sepsis prediction is crucial for timely intervention and improved patient outcomes.
• Existing detection tools often exhibit low performance and depend on time-intensive laboratory results.

Purpose of the Study:

• To validate the InSight sepsis prediction method against Sepsis-3 definitions using retrospective electronic health record data.
• To compare InSight's performance with established scoring systems like qSOFA, MEWS, SIRS, SAPS II, and SOFA.
• To assess the impact of data sparsity on the robustness and accuracy of the InSight system.

Main Methods:

• InSight, a machine learning classifier utilizing vital signs, oxygen saturation, Glasgow Coma Score, and age, was applied to the MIMIC-III dataset.
• Performance was evaluated against Sepsis-3 criteria, comparing InSight with qSOFA, MEWS, SIRS, SAPS II, and SOFA.
• The system's resilience to random data deletion was tested to evaluate performance under data sparsity.

Main Results:

• InSight demonstrated superior classification performance (AUROC=0.880, APR=0.595) compared to all other tested scores at sepsis onset.
• Performance remained significantly higher even when up to 60% of input data was randomly deleted (AUROC=0.781, APR=0.401).
• InSight outperformed SIRS, qSOFA, SAPS II, and SOFA even with substantial data loss, highlighting its robustness.

Conclusions:

• InSight effectively predicts sepsis onset using readily available patient data, primarily vital signs.
• The machine learning approach shows remarkable performance and robustness, even with significant data missingness.
• InSight represents a promising tool for early and accurate sepsis detection in clinical settings.