Validation of a Machine Learning Model for Early Shock Detection

Area of Science:

• Critical Care Medicine
• Machine Learning Applications
• Diagnostic Performance Evaluation

Background:

• Early detection of circulatory shock is crucial for patient outcomes.
• Machine learning (ML) models offer potential for real-time clinical decision support.
• The Trauma Triage, Treatment, and Training Decision Support (4TDS) model was developed for shock detection.

Purpose of the Study:

• To prospectively evaluate the real-time performance of the 4TDS ML model for shock detection.
• To compare the diagnostic accuracy of the 4TDS model against the gold standard of electronic medical records (EMRs) review.
• To assess key performance metrics including sensitivity, specificity, and predictive values.

Main Methods:

• A single-center, prospective diagnostic performance study was conducted.
• The study included adult patients admitted to intensive care units (ICUs) and progressive care units.
• The 4TDS model's alerts were compared to clinician EMR review for shock diagnosis.

Main Results:

• The 4TDS model achieved an area under the receiver operating characteristics curve of 0.86.
• Sensitivity was 78.6% and specificity was 93.1% for shock detection.
• The model demonstrated a negative predictive value of 98.4%.

Conclusions:

• The 4TDS ML model was successfully validated for detecting circulatory shock in an ICU setting.
• The model, utilizing only vital signs, showed moderate performance compared to EMR review.
• This study supports the potential utility of ML-driven decision support tools in critical care.