Utilizing artificial intelligence and cellular population data for timely identification of bacteremia in hospitalized patients.

Background: Bacteremia is a critical condition with high mortality that requires prompt detection to prevent progression to life-threatening sepsis. Traditional diagnostic approaches, such as blood cultures, are time-consuming. This limitation has encouraged the exploration of rapid prediction methodologies. Cellular Population Data (CPD), which provides detailed insights into white blood cell morphology and functionality, is a promising technique for the early detection of bacteremia.

Methods: This study applied machine learning models to analyze laboratory data from hospitalized patients at risk of bacteremia from three hospitals. Using complete blood count (CBC), differential count (DC), and CPD, collected at various time intervals, we trained two sets of artificial intelligence models: one trained using data from patients in the Emergency Department (ED) and another specifically designed for and trained using data from a hospitalized cohort. We evaluated the performance of both models by applying them to the same hospitalized population and comparing their outcomes.

Results: The study encompassed analysis of over 66,000 CBC samples. The model tailored for hospitalized patients exhibited superior performance in bacteremia prediction across all cohorts compared with the ED-model, achieving an area under the receiver operating characteristic curve (AUROC) of 0.772 in the validation cohort from China Medical University Hospital and 0.808 and 0.843 in two other hospital cohorts. Notably, nearly half of the top fifteen important features identified by shapely additive explanations values were CPD parameters, underscoring the pivotal role of CPD in predictive models for bacteremia.

Conclusions: Artificial intelligence models incorporating CPD data can accurately predict bacteremia in hospitalized patients. Models specifically trained on hospitalized patient data demonstrate enhanced performance over those based on ED data in predicting bacteremia occurrences. Future research must explore the clinical effects of these models, focusing on their potential to assist physicians in managing antibiotic use and patient health.