We performed a prospective observational cohort study of consecutive adult patients (age ≥16 years) presenting to two EDs in Cape Town, South Africa between April and December 2013. The EDs at Groote Schuur Hospital (GSH) and Victoria Hospital (VH) each attend to approximately 40,000 adult patients per year. Both serve urban communities in the City of Cape Town, which had a 19.1 % HIV seroprevalence rate documented in 2010 . Unlike Victoria Hospital, GSH offers a number of specialist services including bone marrow and solid organ transplants.
The decision to perform a blood culture was made independently by the patient’s medical care provider who completed a form with the clinical details available at the time of venesection. Written informed consent was obtained from all patients. The following data was collected for each patient; co-morbidities, the presence of chills, and use of antibiotics in the previous 48 h, temperature, pulse, blood pressure (BP), respiratory rate, oxygen saturation, inspired oxygen concentration and presence of confusion. The suspicion of a focal infection (source) was determined by the attending clinician based on clinical evaluation and was classified into the following categories: lower respiratory tract (pneumonia, bronchitis or acute exacerbation of chronic obstructive pulmonary disease), urinary tract (lower or upper), endocarditis, skin and soft tissue (cellulitis, wet gangrene, necrotising fasciitis, and abscess), biliary tract, gastrointestinal, gynaecological system, meningitis, source of infection unclear and ‘other’.
Hyperthermia was defined as temperature ≥38.3 °C and hypothermia as temperature <36.0 °C. Tachycardia was defined as pulse rate ≥120 beats/min. CRB-65 score was calculated as 1 point for each of; confusion, high respiratory rate ≥30 per minute, low systolic BP < 90 mm Hg or diastolic BP < 60 mmHg, and advanced age ≥65 years . CRB-65 is a validated predictor of mortality from community acquired pneumonia and systemic infections.
A blood culture ‘set’ consisted of a single aerobic bottle. All bottles were weighed after inoculation of blood to determine the volume added. Standard laboratory procedures according to the National Health Laboratory Service (NHLS) in South Africa were used to identify organisms and determine resistance patterns.
The clinical records of all patients with positive blood cultures were reviewed by two infectious disease specialists (TB, MM) to determine whether the organisms represented true pathogens or contaminants. Disagreements were to be adjudicated by a third Infectious Diseases specialist but there were none. Decisions were based on the discretion of the physicians based on the clinical findings in each individual case. Contaminants were common skin commensals (e.g. coagulase-negative staphylococci, “diphtheroids”, micrococcus, Bacillus spp.), which when identified were not in keeping with the clinical features of the case. Clinical records were reviewed to determine the ability of the result to influence patient management. This was assessed by determining if the isolate was grown from another sterile site, and whether the antimicrobials prescribed by the attending clinician were appropriate for the organism.
Frequency (percentage) and median (inter quartile range [IQR]) were calculated for the entire data set as well as for the groups true bacteraemia versus contaminant or no growth. Simple comparisons between groups were done with Fisher’s exact test for discrete variables and Wilcoxon rank sum test for continuous variables. Regression models with a single explanatory variable were carried out to determine the association between potential diagnostic variables and the outcome BSI on the complete data set for the purpose of data exploration and understanding. A number of approaches were explored to find a clinical prediction rule, modelling was done on a randomly selected development sub-sample (50 % of data), including multivariable regression modelling and classification tree analysis. Full statistical methods details available in the online supplement. A clinical prediction rule was considered if the resulting model performance was sufficient (high negative predictive value, NPV) on the validation set, however, no models performed well enough to consider developing a clinical prediction rule. Visual inspection of deviance statistics and partial residual plots were used to check regression assumptions.
Ethical approval for this study was granted by the Faculty of Health Sciences Human Research Ethics Committee, University of Cape Town. Reference 172/2013.