Development and Evaluation of a Machine Learning Model for the Early Identification of Patients at Risk for Sepsis

Area of Science:

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

Background:

• The quick Sequential [Sepsis-related] Organ Failure Assessment (qSOFA) score is recommended for sepsis screening outside intensive care units (ICUs).
• Concerns exist regarding the sensitivity of qSOFA for early sepsis detection.
• There is a need for improved sepsis screening tools in emergency departments.

Purpose of the Study:

• To develop a novel sepsis screening tool, the Risk of Sepsis (RoS) score, using machine learning.
• To compare the performance of the RoS score against established sepsis screening benchmarks.
• To evaluate the timeliness and discriminant capabilities of the RoS score.

Main Methods:

• Retrospective analysis of electronic health records from 2,759,529 adult patients across 49 urban emergency departments.
• Development of the RoS score using gradient boosting machine learning on a training cohort.
• Evaluation of the RoS score against Systemic Inflammatory Response Syndrome, SOFA, qSOFA, MEWS, and NEWS using key performance metrics.

Main Results:

• The RoS score exhibited superior discriminant ability (AUC 0.93-0.97) compared to all benchmark tools.
• RoS demonstrated significantly higher sensitivity and precision than Sequential Organ Failure Assessment (SOFA) at all time thresholds.
• The quick Sequential [Sepsis-related] Organ Failure Assessment (qSOFA) score showed low sensitivity in this cohort.

Conclusions:

• The developed Risk of Sepsis (RoS) score is a more timely and discriminant tool than current benchmark sepsis screening methods.
• The RoS score shows promise for improving sepsis detection in emergency department settings.
• Further external validation studies are required to confirm the generalizability of the RoS score.