Amid a debilitating nationwide blood culture bottle shortage, hospitals across the US faced a severe resource crunch with some institutions having less than 12 bottles in stock at any given time. Nicholas P. Marshall, MD, FAAP, a Pediatric Infectious Diseases and Clinical Informatics Fellow at Stanford University, saw this as an opportunity to innovate, turning to machine learning with the intent of reducing blood culture bottle usage without compromising patient safety. By way of solution, Marshall and his team instituted a set of open data models capable of predicting the likelihood of a patient displaying bacteremia and determining when a culture could be safely deferred. The clever restructuring of these data models enabled the team to optimize the use of meager resources amid the critical shortage.
Using an extensive cohort derived from Stanford Hospital and Stanford Health Care Tri-Valley and spanning over eight years, the team built models based on more than 135,000 blood cultures. These models varied in complexity from the basic logistic regression model, which utilizes common emergency department labs and vital signs to predict the possibility of bloodstream infections, to the more advanced multimodal algorithm. This algorithm employs a large language model capable of parsing unstructured provider notes and capturing clinical details otherwise absent in structured data.
The team’s efforts were fruitful, yielding models whose performance improved consistently with each more complex level. Their most considerable success came with the multimodal algorithm, which outshone traditional predictive rules for bacteremia like Systemic Inflammatory Response Syndrome and Shapiro criteria.
In the face of a shortage, these models offer invaluable assistance in deciding when to order cultures and when it is safe to defer based on the predicted pre-test probability. Significantly, these tools can reduce the occurrence of false positives, contamination, unnecessary antibiotic prescriptions, and subsequent costs due to their high precision and accuracy. Moreover, these models, designed for widespread adoption, require no additional software or special resources. The team’s innovation not only provides immediate solutions but also has the potential to reshape standard operating procedures beyond periods of scarcity. Marshall and his team have thus delivered an invaluable tool that stands to transform the landscape of infection control and prevention.