Online Sepsis Prediction Using Vital Signs and Multiscale Temporal-Aware Contrastive Learning: Model Development and...

Area of Science:

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

Background:

• * Real-time sepsis prediction is crucial but limited by reliance on delayed laboratory results.
• * Current models struggle with the time-series nature of patient data and are inefficient.
• * Existing methods often require complex data not readily available for timely diagnosis.

Purpose of the Study:

• * To develop an online sepsis detection model using only easily accessible vital signs.
• * To leverage multiscale temporal representation learning for variable-length input sequences.
• * To maintain high predictive performance in real-time sepsis prediction.

Main Methods:

• * Proposed a deep learning model: Multi-Scale Temporal-aware Contrastive Learning (MSTCL).
• * Utilized multiscale temporal modeling to capture short- and long-term dependencies in physiological time series.
• * Employed contrastive learning for robustness, differentiating sepsis progression trajectories using 6 vital signs.

Main Results:

• * Achieved 88.34% AUC, 89.29% sensitivity, and 73% specificity in predicting sepsis onset.
• * Demonstrated high performance on over 400 patients using variable-length vital-sign histories.
• * Reported a normalized mean absolute error of 0.11% for predicted sepsis onset.

Conclusions:

• * MSTCL offers low complexity and rapid inference, suitable for real-time monitoring systems.
• * The model's ability to learn from variable-length data enhances clinical applicability.
• * Temporal-aware contrastive learning provides a robust solution for online sepsis detection in various clinical settings.