Practical Machine Learning-Based Sepsis Prediction.

Area of Science:

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

Background:

• Sepsis is a life-threatening condition with nonspecific symptoms, complicating early diagnosis.
• Current sepsis detection relies on clinical expertise and laboratory biomarkers, often requiring intensive care.
• Improving sepsis recognition and treatment is vital for patient outcomes and reducing healthcare costs.

Purpose of the Study:

• To evaluate the impact of different feature sets on machine learning model performance for sepsis prediction.
• To assess the efficacy of vital sign measures versus expert and biomarker data.
• To explore the role of population homogeneity in sepsis prediction precision.

Main Methods:

• Utilized XGBoost machine learning models for sepsis prediction.
• Trained models on heterogeneous (n=3932) and homogeneous (n=1927) intensive care unit (ICU) patient cohorts.
• Systematically removed clinical expert and laboratory biomarker features to isolate the impact of vital signs.

Main Results:

• Removing expert and biomarker features significantly degraded model performance (p<0.05) in heterogeneous ICU populations.
• Models trained solely on vital signs achieved significantly improved moderate prediction (AUPRC, P<0.05) in homogeneous ICU patient groups.
• Feature importance varied based on population characteristics and data inclusion.

Conclusions:

• Vital sign data alone can achieve moderate sepsis prediction, particularly in homogeneous patient groups.
• Practical machine learning systems for sepsis prediction should consider feature availability and population demographics.
• Personalized sepsis prediction models tailored to specific patient populations are essential for improved clinical utility.