A rapid and accurate detection of carbapenemase-producing Gram-negative bacteria (CPGNB) has an immediate demand in the clinic. Here, we developed and validated a method for rapid detection of CPGNB using Blue-Carba combined with deep learning (designated as AI-Blue-Carba). The optimum bacterial suspension concentration and detection wavelength were determined using a Multimode Plate Reader and integrated with deep learning modeling. We examined 160 carbapenemase-producing and non-carbapenemase-producing bacteria using the Blue-Carba test and a series of time and optical density values were obtained to build and validate the machine models. Subsequently, a simplified model was re-evaluated by descending the dataset from 13 time points to 2 time points. The best suitable bacterial concentration was determined to be 1.5 optical density (OD) and the optimum detection wavelength for AI-Blue-Carba was set as 615 nm. Among the 2 models (LRM and LSTM), the LSTM model generated the higher ROC-AUC value. Moreover, the simplified LSTM model trained by short time points (0-15 min) did not impair the accuracy of LSTM model. Compared with the traditional Blue-Carba, the AI-Blue-Carba method has a sensitivity of 95.3% and a specificity of 95.7% at 15 min, which is a rapid and accurate method to detect CPGNB.
Keywords: Blue-Carba; OD value; carbapenemase-producing gram-negative bacteria; deep learning; rapid detection.
Copyright © 2020 Jia, Han, Cai, Cui, Yang, Zhang, Bai, Liu, Wei, Chen, Liao, Liu, Li and Sun.