ABSTRACT
Accurate and early prediction of poststroke infections is important to improve antibiotic therapy guidance and/or to avoid unnecessary antibiotic treatment. We hypothesized that the combination of blood biomarkers with clinical parameters could help to optimize risk stratification during hospitalization. In this prospective observational study, blood samples of 283 ischemic stroke patients were collected at hospital admission within 72 h from symptom onset. Among the 283 included patients, 60 developed an infection during the first five days of hospitalization. Performance predictions of blood biomarkers (Serum Amyloid-A (SAA), C-reactive protein, procalcitonin (CRP), white blood cells (WBC), creatinine) and clinical parameters (National Institutes of Health Stroke Scale (NIHSS), age, temperature) for the detection of poststroke infection were evaluated individually using receiver operating characteristics curves. Three machine learning techniques were used for creating panels: Associative Rules Mining, Decision Trees and an internal iterative-threshold based method called PanelomiX. The PanelomiX algorithm showed stable performance when applied to two representative subgroups obtained as splits of the main subgroup. The panel including SAA, WBC and NIHSS had a sensitivity of 97% and a specificity of 45% to identify patients who did not develop an infection. Therefore, it could be used at hospital admission to avoid unnecessary antibiotic (AB) treatment in around half of the patients, and consequently, to reduce AB resistance.
