Abstract
Staphylococcus epidermidis is an important cause of nosocomial infection and bacteremia. It is also a common contaminant of blood cultures and, as a result, there is frequently uncertainty as to its diagnostic significance when recovered in the clinical laboratory. One molecular strategy that might be of value in clarifying the interpretation of S. epidermidis identified in blood culture is multilocus sequence typing. Here, we examined 100 isolates of this species (50 blood isolates representing true bacteremia, 25 likely contaminant isolates, and 25 skin isolates) and the ability of sequence typing to differentiate them. Three machine learning algorithms (classification regression tree, support vector machine, and nearest neighbor) were employed. Genetic variability was substantial between isolates, with 44 sequence types found in 100 isolates. Sequence types 2 and 5 were most commonly identified. However, among the classification algorithms we employed, none were effective, with CART and SVM both yielding only 73% diagnostic accuracy and nearest neighbor analysis yielding only 53% accuracy. Our data mirror previous studies examining the presence or absence of pathogenic genes in that the overlap between truly significant organisms and contaminants appears to prevent the use of MLST in the clarification of blood cultures recovering S. epidermidis.