Children are small adults (when properly normalized): Transferrable/generalizable sepsis prediction

Area of Science:

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

Background:

• Pediatric sepsis presents unique physiological trajectories distinct from adult cases.
• Limited pediatric sepsis datasets hinder direct model development.
• Leveraging shared biology requires normalizing pediatric data for adult comparability.

Purpose of the Study:

• To normalize pediatric physiological data for direct comparison with adult data.
• To develop machine learning classifiers for predicting pediatric sepsis onset.
• To externally validate these classifiers on an independent adult dataset.

Main Methods:

• Utilized vital signs and laboratory data from the Pediatric Intensive Care (PIC) database.
• Developed the Continuous Age-Normalized SOFA (CAN-SOFA) score for age normalization.
• Employed the XGBoost algorithm for sepsis classification and validated on MIMIC-IV adult data.

Main Results:

• The pediatric sepsis classifier achieved 0.84 accuracy and 0.867 F1-Score.
• On adult data, the classifier showed 0.80 accuracy and 0.88 F1-Score.
• Observed similar performance degradation (data drift) in both populations when tested externally.

Conclusions:

• Demonstrated the generalizability of Electronic Health Records (EHRs) between pediatric and adult populations for sepsis prediction.
• A straightforward age-normalization method enables cross-population applicability.
• This approach facilitates leveraging shared biological underpinnings for improved sepsis detection.