Artificial intelligence-based gram stain classification: Accuracy and clinical utility in positive blood cultures.

IntroductionWith the rapid advancement of artificial intelligence (AI) technology, AI has been applied to the detection of pathogens that cause infectious diseases. This study evaluated the newly developed AI-based software, the BiTTE-iE, based on Gram-stained specimen images captured with smartphones, assessing not only its ability to accurately identify bacterial species but also its capability to classify pathogens morphologically.MethodsGram-stained slides were prepared from 177 positive blood culture bottles as part of routine clinical practice. This study evaluated the match rate between the results of "Major Pathogen Classification" (GPCs, GPRs, GNCs, GNRs, and Yeast) performed by the BiTTE-iE and the microscopic findings of microbiology-trained medical technologists. Furthermore, the match rate between the pathogen classifications identified by the BiTTE-iE at the "Group-level" and "Species-level" and the final species identification results obtained using culture methods and MALDI-TOF MS was evaluated.ResultsOverall, the match rate for the "Major Pathogen Classification" exceeded 85% in all classification categories except GNCs, where data collection was impossible. For GPCs, a high match rate of 93.8% (166/177) was observed. In contrast, the match rate for "Species-level" classification was relatively low, generally ranging from 0% to 80%. In particular, the match rates for Enterococcus spp. and non-fermentative Gram-negative rods were especially low, resulting in an overall "Species-level" match rate of 41.5% (66/159).ConclusionsThe BiTTE-iE demonstrated potential utility in selecting antimicrobial agents for empirical treatment. By comprehensively accumulating image data for all microbials, it is expected that diagnostic performance using the BiTTE-iE will be further improved.