Gram staining decipherment using an artificial intelligence-powered smartphone-based application.

Gram staining provides rapid microbiological information that may assist in empirical antimicrobial selection; however, the results are often interpreted by microbiological specialists who are not always available. Therefore, we developed a computer-aided diagnosis system using artificial intelligence trained on microscopic images of Gram-stained urine, captured with an iPhone, using the Bartholomew and Mittwer method. The system interprets Gram-stained urine samples and classifies bacterial morphology (Class 1: 7 predefined morphology categories) and 17 predefined species-level categories (Class 2). In this retrospective observational study, five imaging devices and two staining methods (Bartholomew and Mittwer, Favor) were compared. Urine specimens were collected from two hospitals between 1 April and 31 December 2022. Validation images were generated using five devices (four smartphones and one microscopic camera). We used a micrometer with microscopy with all smartphones; some iPhone images were taken without a micrometer. Favor staining was only imaged using an iPhone without the micrometer. Image data sets were generated from 433 clinical and 17 spiked samples. The overall accuracy was 0.804 for Class 1 and 0.640 for Class 2. Images taken by the microscopic camera had the highest accuracy and kappa coefficient, whereas the AQUOS smartphone had the lowest accuracy and kappa coefficient. The accuracy of images created without a micrometer was 0.885 for Class 1 and 0.666 for Class 2. The Bartholomew and Mittwer method had better accuracy and a better kappa coefficient. Overall, accuracy depended on the staining method used in the training data, not on the imaging device.IMPORTANCEGram staining provides rapid information on both the site of infection and likely pathogens, guiding empirical antimicrobial selection. However, interpretation requires infectious disease expertise, which is not always available. We developed an artificial intelligence-based diagnostic support system trained on iPhone images of Gram-stained urine using the Bartholomew and Mittwer method to classify bacterial morphology (Class 1) and inferred species (Class 2). To provide essential baseline data on factors influencing accuracy and reliability, we compared Gram-stained urine images from two hospitals obtained with five imaging devices and two staining methods. Microscopic camera images showed the highest accuracy, whereas an AQUOS smartphone showed the lowest. Images without a micrometer performed better, and the Bartholomew and Mittwer method outperformed the Favor method. Accuracy increased when confidence levels were higher. Our findings suggest that using the same staining method as the training data and avoiding micrometer noise are critical, while device differences are less influential.