Longitudinal Biomarker Trajectories for EVD-Associated Infection Prediction: A Trajectory-Based Machine Learning...

Area of Science:

• Neurocritical care
• Infectious disease prediction
• Machine learning in medicine

Background:

• External ventricular drain (EVD)-associated infections occur in 1-40% of patients.
• Assessing cerebrospinal fluid (CSF) parameters is challenging, particularly in hemorrhagic neurological disease.
• This study addresses the need for real-time EVD-associated infection prediction.

Purpose of the Study:

• To develop and validate a machine learning framework for real-time prediction of EVD-associated infection.
• To identify key biomarkers and their temporal dynamics predictive of infection.
• To determine the lead time between biomarker deterioration and clinical diagnosis.

Main Methods:

• Retrospective analysis of 367 neurocritical patients with EVDs (8,419 patient-days).
• Utilized seven serially monitored biomarkers: WBC, CRP, procalcitonin, CSF cell count, lactate, glucose, and protein.
• Developed a machine learning ensemble and a composite scoring algorithm for biological onset analysis.

Main Results:

• 88 out of 367 patients (24%) developed EVD-associated infection.
• The machine learning ensemble achieved an AUC-ROC of 0.833 at 24 hours and 0.822 at 48 hours.
• Biomarker deterioration preceded clinical diagnosis by a median of 2 days, with dynamic WBC and CSF lactate being key predictors.

Conclusions:

• Longitudinal biomarker monitoring combined with machine learning enables timely prediction of EVD-associated infection.
• Biomarker changes offer a lead time of approximately two days over clinical recognition.
• Supports prospective evaluation of daily risk monitoring for EVD-associated infections in neurocritical care.