A customised down-sampling machine learning approach for sepsis prediction.

Area of Science:

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

Background:

• Sepsis is a critical ICU condition requiring timely intervention.
• Existing sepsis prediction models face challenges with alarm fatigue and patient safety.
• Accurate, robust, and deployable prediction methods using vital signs and lab tests are needed.

Purpose of the Study:

• To develop an accurate and robust sepsis prediction method for ICUs.
• To utilize only vital signs and laboratory tests for sepsis prediction.
• To mitigate alarm fatigue by improving prediction specificity.

Main Methods:

• A customized down-sampling process was applied to retrospective data for XGBoost model training.
• A dynamic sliding window approach was integrated with the trained XGBoost model.
• The method was evaluated on the PhysioNet (PhysioNet-A, PhysioNet-B) and FHC retrospective datasets.

Main Results:

• The method achieved 80.74% accuracy (77.90% sensitivity, 84.42% specificity) on PhysioNet-A and 83.95% accuracy (84.82% sensitivity, 82.00% specificity) on PhysioNet-B.
• AUC scores of 0.89 were obtained for both PhysioNet datasets.
• On the FHC dataset, 92.38% accuracy (88.37% sensitivity, 95.16% specificity) and an AUC score of 0.98 were achieved.

Conclusions:

• The developed method demonstrates robust and accurate sepsis prediction capabilities across different ICU settings.
• The combination of down-sampling, dynamic sliding window, and XGBoost effectively predicts sepsis.
• This localized and robust method can aid sepsis diagnosis in diverse ICU environments.