With an estimated 440,000 active cases occurring each year, medical device associated infections pose a significant burden on the US healthcare system, costing about $9.8 billion in 2013. Staphylococcus epidermidis is the most common cause of these device-associated infections, which typically involve isolates that are multi-drug resistant and possess multiple virulence factors. S. epidermidis is also frequently a benign contaminant of otherwise sterile blood cultures. Therefore, tests that distinguish pathogenic from non-pathogenic isolates would improve the accuracy of diagnosis and prevent overuse/misuse of antibiotics. Attempts to use multi-locus sequence typing (MLST) with machine learning for this purpose had poor accuracy (~73%). In this study we sought to improve the diagnostic accuracy of predicting pathogenicity by focusing on phenotypic markers (i.e., antibiotic resistance, growth fitness in human plasma, and biofilm forming capacity) and the presence of specific virulence genes (i.e., mecA, ses1, and sdrF). Commensal isolates from healthy individuals (n = 23), blood culture contaminants (n = 21), and pathogenic isolates considered true bacteremia (n = 54) were used. Multiple machine learning approaches were applied to characterize strains as pathogenic vs non-pathogenic. The combination of phenotypic markers and virulence genes improved the diagnostic accuracy to 82.4% (sensitivity: 84.9% and specificity: 80.9%). Oxacillin resistance was the most important variable followed by growth rate in plasma. This work shows promise for the addition of phenotypic testing in clinical diagnostic applications.

With an estimated 440,000 active cases occurring each year, medical device associated infections pose a significant burden on the US healthcare system, costing about $9.8 billion in 2013. Staphylococcus epidermidis is the most common cause of these device-associated infections, which typically involve isolates that are multi-drug resistant and possess multiple virulence factors. S. epidermidis is also frequently a benign contaminant of otherwise sterile blood cultures. Therefore, tests that distinguish pathogenic from non-pathogenic isolates would improve the accuracy of diagnosis and prevent overuse/misuse of antibiotics. Attempts to use multi-locus sequence typing (MLST) with machine learning for this purpose had poor accuracy (~73%). In this study we sought to improve the diagnostic accuracy of predicting pathogenicity by focusing on phenotypic markers (i.e., antibiotic resistance, growth fitness in human plasma, and biofilm forming capacity) and the presence of specific virulence genes (i.e., mecA, ses1, and sdrF). Commensal isolates from healthy individuals (n=23), blood culture contaminants (n=21), and pathogenic isolates considered true bacteremia (n=54) were used. Multiple machine learning approaches were applied to characterize strains as pathogenic vs non-pathogenic. The combination of phenotypic markers and virulence genes improved the diagnostic accuracy to 82.4% (sensitivity: 84.9% and specificity: 80.9%). Oxacillin resistance was the most important variable followed by growth rate in plasma. This work shows promise for the addition of phenotypic testing in clinical diagnostic applications.