Hospital-Wide Sepsis Detection: A Machine Learning Model Based on Prospectively Expert-Validated Cohort

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support Systems

Background:

• Sepsis detection is challenging due to clinical variability and limitations of current scoring systems.
• Traditional methods often lack the sensitivity and specificity required for timely diagnosis.
• Developing advanced diagnostic tools is crucial for improving patient outcomes.

Purpose of the Study:

• To develop and validate a hospital-wide machine learning model for enhanced sepsis detection.
• To improve diagnostic accuracy beyond conventional scoring systems.
• To leverage both structured and unstructured clinical data for a comprehensive model.

Main Methods:

• Retrospective analysis of 218,715 hospital episodes (2014-2018).
• Integration of structured data (26.95%) and unstructured clinical notes (73.04%) via natural language processing (NLP).
• Development and evaluation of 30 machine learning models, with the best performing model (BiAlert Sepsis) selected.

Main Results:

• The BiAlert Sepsis model achieved an AUC-ROC of 0.95, sensitivity of 0.93, and specificity of 0.84.
• Significantly outperformed traditional approaches, reducing false positives by 39.6% compared to Sepsis-2 + qSOFA.
• Identified novel predictors like eosinopenia and hypoalbuminemia, alongside traditional variables.

Conclusions:

• The hospital-wide machine learning model demonstrates superior sepsis detection performance.
• Integration of NLP-derived features from expert-validated cases enhances diagnostic capabilities.
• Further external validation and prospective studies are recommended before widespread implementation.