Predicting sepsis using deep learning across international sites: a retrospective development and validation study.

Area of Science:

• Artificial Intelligence in Medicine
• Critical Care Medicine
• Machine Learning for Healthcare

Background:

• Early sepsis detection is crucial as organ damage may be irreversible upon diagnosis.
• Machine learning shows promise for early sepsis prediction, but lacks international validation.
• Current diagnostic methods often lag behind disease progression, impacting patient prognosis.

Purpose of the Study:

• To develop and externally validate a deep learning system for early sepsis prediction in intensive care units (ICUs).
• To assess the generalizability of the deep learning model across diverse international ICU cohorts.
• To compare the model's performance against existing clinical and machine learning baselines.

Main Methods:

• Retrospective, observational, multi-center cohort study involving 136,478 ICU admissions from the US, Netherlands, and Switzerland (2001-2016).
• Development of a deep learning system using hourly-resolved data and Sepsis-3 definition for sepsis annotation.
• Extensive internal and external validation across multiple databases, reporting Area Under the Receiver-Operating Characteristic Curve (AUC).

Main Results:

• The deep learning model achieved an average AUC of 0.846 internally and 0.761 externally across sites.
• Fine-tuning with 10% of site data improved external validation AUC to 0.807.
• The model detected 80% of sepsis cases 3.7 hours prior to onset with a low false alert rate (1.4 per true alert).

Conclusions:

• A deep learning system can generalize internationally for real-time sepsis detection in ICUs.
• The model provides a significant early warning window, enabling timely clinical interventions.
• This study represents the first international, multi-center validation of deep learning for ICU sepsis prediction.