When a nationwide blood culture bottle shortage hit last year, some US hospitals were down to “less than even 12 bottles available” at a time, recalls Nicholas P. Marshall, MD, FAAP, a pediatric infectious diseases and clinical informatics fellow at Stanford University. “We wanted to find a way that we could safely reduce blood culture bottle usage,” he told Infection Control Today® at IDWeek 2025, held in Atlanta, Georgia, from October 19 to 22. The result is a set of open, data-driven models to help clinicians decide when a culture is most likely to be positive—and when it isn’t.
A Large, Shareable Cohort
Marshall’s team assembled an 8-year, 2-institution cohort from Stanford Hospital and Stanford Health Care Tri-Valley, restricting to adults (18 years or older). The final dataset includes more than 135,000 blood cultures and is publicly available in the group’s Antimicrobial Resistance Database. “You can…use it for training other machine learning models as well,” Marshall said.
To mimic real-world deployment and avoid temporal data leakage, the data were split chronologically into training, test, and validation sets.
Three Escalating Models
Logistic regression (baseline): Using vital signs plus approximately 10 common emergency department labs, this model predicts the probability of bloodstream infection (BSI).
Kaltrics Score (bedside calculator): The logistic regression was converted to a point-based tool that any hospital can implement without new software. “If the patient has a fever, that will give you 3 more points,” said Marshall. Age, tachycardia, hypotension, and lactate also influence the score.
Cultrix Integrated (expanded features): This model layers in diagnoses and recent prescriptions, especially prior antibiotic exposure, yielding better discrimination than vitals/labs alone.
Multimodal algorithm (LLM-augmented): The team’s most advanced approach harnesses a large language model to parse unstructured provider notes, capturing clinical nuance that structured data miss. “We created a multimodal algorithm…to look for those unstructured clinical data points,” Marshall explained.
Outperforming Common Rules
Across the board, performance improved as the models grew richer with higher sensitivity and specificity from logistic regression to Cultrix Integrated to the LLM-augmented multimodal approach. Importantly, the models outperformed traditional rules such as Systemic inflammatory response syndrome and the Shapiro criteria for predicting bacteremia.
Practical Value During Scarcity and Beyond
The immediate use case is clear: in a shortage, order cultures when the model indicates high pre-test probability and consider deferring when risk is low, paired, of course, with clinical judgment. But the upside extends beyond shortages. Fewer low-yield cultures can reduce false positives, contamination, unnecessary antibiotics, and downstream costs.
Best of all, the tools were built for broad adoption. “You don’t have to buy anything for your hospital,” Marshall said of the Kaltrics Score, which can be implemented quickly as a calculator in the EHR or on a smartphone. Centers with data science capacity can trial the Integrated or Multimodal versions for even better accuracy.
Why IDWeek Matters
For Marshall, the conference is fuel for collaboration. “The best part…is seeing all of the cool things that are happening with my colleagues,” he said. “You can collaborate together but also be inspired by other specialties… and how your work can help them move forward.”
