The development an artificial intelligence algorithm for early sepsis diagnosis in the intensive care unit

Area of Science:

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

Background:

• Severe sepsis and septic shock are leading causes of death in ICUs, necessitating timely diagnosis.
• Electronic medical records (EMR) offer vast clinical data potential for AI development, but data heterogeneity poses challenges.
• Manual data processing in EMR hinders AI progress in critical care settings.

Purpose of the Study:

• To develop and evaluate an AI algorithm for sepsis diagnosis using pre-selected features from EMR data.
• To compare the diagnostic performance of the AI algorithm against the SOFA score.

Main Methods:

• A prospective, open-label cohort study utilizing a specialized EMR (TED_ICU) for continuous data recording.
• Selection of 106 clinical features relevant to sepsis diagnosis and daily recording.
• Development of an AI algorithm using XGBoost machine learning on de-identified data, with 5-fold cross-validation.

Main Results:

• The AI algorithm achieved 82% accuracy, 65% sensitivity, and 88% specificity, with an AUROC of 0.89.
• The SOFA score-based diagnostic method showed inferior performance with an AUROC of 0.596.
• The AI algorithm demonstrated superior diagnostic capability compared to the SOFA score in the studied ICU cohort.

Conclusions:

• The developed AI algorithm provides timely and accurate sepsis diagnosis using real-time EMR data, exceeding 80% accuracy.
• The AI algorithm outperforms the SOFA score, enabling earlier clinical intervention and improved patient outcomes.
• Early and precise sepsis diagnosis via AI offers significant benefits, including cost reduction and enhanced healthcare system efficiency.