An Improved Patient-Specific Mortality Risk Prediction in ICU in a Random Forest Classification Framework

Area of Science:

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

Background:

• Timely prediction of Intensive Care Unit (ICU) mortality is crucial for effective clinical intervention.
• Existing mortality prediction tools may lack the dynamic, patient-specific accuracy needed in critical care settings.

Purpose of the Study:

• To design and validate a novel patient-specific mortality prediction system for ICUs.
• To leverage time-series vital signs and laboratory data for enhanced risk stratification.
• To improve upon the performance of established ICU mortality prediction algorithms.

Main Methods:

• Utilized a dataset of 4000 adult ICU patients, incorporating demographic information and time-series measurements from the first 48 hours.
• Developed an ensemble of decision trees for simultaneous risk score prediction and association.
• Employed a k-fold cross-validation framework for robust model evaluation.

Main Results:

• Achieved a risk assessment prediction accuracy of 87% for ICU mortality.
• Demonstrated significant performance improvement compared to the commonly used SAPS-I baseline algorithm.
• Results were further benchmarked against other state-of-the-art algorithms on the same dataset.

Conclusions:

• The developed ensemble model offers a highly accurate and dynamic approach to ICU mortality prediction.
• This system has the potential to facilitate timely and appropriate interventions by healthcare professionals.
• The findings suggest a promising advancement in machine learning applications for critical care decision support.