On classifying sepsis heterogeneity in the ICU: insight using machine learning.

Area of Science:

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

Background:

• Current machine learning models for sepsis prediction from electronic health records (EHR) overlook patient heterogeneity.
• Sepsis heterogeneity significantly impacts prognosis and treatment, necessitating more nuanced predictive approaches.

Purpose of the Study:

• To demonstrate the value of stratifying sepsis patients based on organ dysfunction types.
• To improve the accuracy of recognizing patients at risk of sepsis using EHR data.

Main Methods:

• Analysis of a large intensive care unit (ICU) dataset (13,728 records).
• Identification of distinct sepsis subpopulations based on organ dysfunction patterns.
• Classification experiments using Random Forest, Gradient Boost Trees, and Support Vector Machines.

Main Results:

• Sepsis subpopulations with distinct organ dysfunction patterns were identified.
• Features selected using sepsis subpopulations improved classification performance.
• Enhanced specificity was observed in distinguishing septic from non-septic patients, overcoming a current bottleneck.

Conclusions:

• Stratifying patients by organ dysfunction enhances machine learning model performance for sepsis prediction.
• Findings support the development of more personalized machine learning models for complex conditions like sepsis.