A model for predicting bacteremia species based on host immune response

Tione Buranda ^1* Peter Simons ¹ Virginie Bondu ¹ Laura Shevy ¹ Stephen Young ² Angela Wandinger-Ness ¹ Cristian G Bologa ¹

• ¹ University of New Mexico, Albuquerque, United States
• ² TriCore Reference Laboratories (United States), Albuquerque, New Mexico, United States

Clinicians face challenges in rapidly and accurately identifying the bacterial species that cause patient bacteremia and selecting appropriate antibiotics for timely treatment. In this study, we implemented a novel approach to address these issues by combining immune response data obtained from routine blood counts with assessments of immune cell activation, specifically through quantitative measurements of Rho family GTPase activity. The combined data were then used to develop a machine-learning model that distinguishes specific classes of bacteria and their memberships. We specifically investigated whether different classes of bacteria provoke distinct patterns of host immune responses, as indicated by quantitative differences in leukocyte populations derived from routine complete blood counts with differential. In parallel, quantitative measurements of activated Rac1 (Rac1•GTP) levels were conducted using a novel 'G-Trap assay' we developed. Using G-Trap, we measured Rac1•GTP in peripheral blood monocytes (PBMC) and polymorphonuclear (PMN) cells from peripheral blood samples collected from 28 culture-positive patients and over 80 non-infected patients used as controls. The results indicated that 18 of the 28 patients with bacteremia exhibited a ≥ 3-fold increase in Rac1•GTP compared to the controls. The remaining ten patients with bacteremia had neutrophilia or pancytopenia and displayed normal to below-normal Rac1 GTPase activity, consistent with bacteria-induced immunosuppression. To analyze the data, we employed partial least squares discriminant analysis (PLS-DA), a supervised method that optimizes group separation and facilitates the development of a novel machine-learning model for pathogen identification. The results demonstrated that PLS-DA effectively differentiates between specific pathogen groups, and an external validation test confirmed the utility of the predictive model. Given that bacterial culture confirmation can take several days, our study highlights that combining routine assays with a machine-learning model can serve as a valuable clinical decision-support tool, enabling prompt and accurate treatment on the same day patients present to the clinic.