Clinical performance of the automated gram stain AI system Mycrium for positive blood culture.

BACKGROUND

Gram staining of positive blood culture bottles provides rapid information for antimicrobial decision-making. This study evaluated the clinical performance and discordance profile of Mycrium, an automated Gram staining and artificial intelligence-based image analysis system, for positive blood culture specimens.

METHODS

We evaluated 231 positive blood culture bottles at Saitama Medical University Hospital. After excluding 2 culture-negative false-positive bottles, 229 culture-positive bottles (208 monomicrobial and 21 polymicrobial) were included in the analysis. Mycrium outputs for Gram stain categories-Gram-positive cocci (GPC), Gram-positive rods (GPR), Gram-negative rods (GNR), and yeast-were compared with routine microscopy performed by a medical technologist (MT) and culture-based organism identification.

RESULTS

Among the monomicrobial bottles, category-level sensitivity was 97.2% for GPC, 94.4% for GNR, 63.2% for GPR, and 60.0% for yeast. Specificity was 86.0% for GPC, 98.5% for GNR, 97.4% for GPR, and 94.9% for yeast. Among the polymicrobial bottles, complete category agreement without additional false-positive detections was achieved in 13/21 bottles (61.9%). Notably, Mycrium identified reference organism categories that were not reported by routine MT microscopy in 4 polymicrobial bottles.

CONCLUSIONS

Mycrium demonstrated high performance for GPC and GNR detection and may facilitate recognition of coexisting organism categories in polymicrobial positive blood cultures. However, the additional false-positive detections and lower sensitivity for GPR and yeast suggest that, at present, Mycrium should be used as a quality-support tool under MT review rather than a standalone replacement for expert microscopy.