Real-time mortality prediction in the Intensive Care Unit

Area of Science:

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

Background:

• Real-time mortality prediction in intensive care units (ICUs) can aid clinical decision-making.
• Synthesizing patient acuity through interpretable models is crucial for effective care.

Purpose of the Study:

• To develop and evaluate a machine learning model for predicting ICU patient mortality.
• To assess the model's performance using data sampled randomly during ICU stays.
• To compare the model's predictive power against established severity of illness scores.

Main Methods:

• Extracted data from a random time point during each patient's ICU stay.
• Developed a Gradient Boosting model for mortality prediction.
• Compared model performance (AUROC) against the Simplified Acute Physiology Score II (SAPS II) using data from the first 24 hours.

Main Results:

• The Gradient Boosting model demonstrated high predictive accuracy (AUROC = 0.920) without using diagnosis or comorbidity data.
• The model significantly outperformed SAPS II (AUROC = 0.927 vs. 0.809) when using data from the initial 24 hours of ICU stay.

Conclusions:

• A Gradient Boosting model offers a highly accurate and interpretable method for real-time mortality prediction in ICUs.
• This approach, utilizing randomly sampled data, shows potential for broad application across a patient's entire ICU trajectory.
• The model's superior performance over SAPS II suggests a promising advancement in critical care patient assessment.