Cultryx: Precision Diagnostic Stewardship for Blood Cultures Using Machine Learning

Area of Science:

• Clinical diagnostics
• Machine learning in healthcare
• Infectious disease management

Background:

• The 2024 blood culture bottle shortage highlighted issues in diagnostic resource allocation.
• Persistent challenges exist with low-value testing and empiric treatment under clinical uncertainty.

Purpose of the Study:

• To evaluate a machine learning (ML) approach using electronic medical record data for bacteremia prediction.
• To determine if ML can guide diagnostic testing and empiric treatment more effectively than current practices.

Main Methods:

• A retrospective cohort of 101,812 adult emergency department encounters (2015-2025) was analyzed.
• An XGBoost model (Cultryx) was trained to predict bacteremia.
• Performance was benchmarked against clinical heuristics (SIRS, Shapiro Rule) and an idealized cognitive baseline (physicians and GPT-5 using Fabre framework).

Main Results:

• Cultryx (AUROC 0.810) outperformed clinical heuristics; SIRS lacked specificity (41.2%), Shapiro Rule lacked sensitivity (70.2%).
• Physicians achieved 95.7% sensitivity with the Fabre framework, while GPT-5 achieved 71.6%.
• Cultryx achieved a 26.2% culture volume deferral rate (deferring ~15,872 bottles) with 98.9% negative predictive value, and Cultryx score retained a 20.8% deferral rate.

Conclusions:

• Machine learning offers a data-driven alternative to clinical heuristics for bacteremia prediction.
• Cultryx can conserve diagnostic resources and reduce unnecessary antibiotic exposure by maximizing culture deferment.
• This approach enhances patient safety by improving pathogen detection and reducing empiric antibiotic use.