Smartphone-based multispectral autofluorescence analysis of bacteria mixtures of staphylococci using convolutional neural network.

Despite the numerous merits of bioreceptor-based biosensors for detecting bacteria, they often fail to identify complex bacterial mixtures, such as those encountered in microbiome analysis, especially when the bacterial mixtures belong to the same genus. Additionally, their procedure involves multiple steps of reagent rinsing and/or staining. This work demonstrates the proof-of-concept for smartphone-based multispectral autofluorescence analysis of bacterial mixtures belonging to the same genus: Staphylococcus epidermidis, Staphylococcus haemolyticus, Staphylococcus capitis, and Staphylococcus aureus, that does not require bioreceptors or staining processes. These staphylococci mixture models represent the skin infection model, where the goal is to identify the presence of pathogenic S. aureus (coagulase-positive) from other Staphylococcus spp. (coagulase-negative staphylococci or CoNS). We designed and tested a portable, inexpensive smartphone-based imaging platform using a smartphone microscope attachment with a 3D-printed housing, LEDs for sample excitation, and low-cost color films for filtering fluorescent emission. For each data point, the excitation light sources and emission filters were alternated by sliding the color films into place and pressing the LED buttons to acquire nine unique autofluorescence images. A convolutional neural network (CNN) model was developed for various bacterial mixtures of CoNS and CoNS + S. aureus. It demonstrated excellent performance in detecting S. aureus presence, with a sensitivity of 91%. However, a small number of false positives compromised the overall accuracy to 84%. Experiments were further replicated with the CoNS- and S. aureus-spiked human skin swab samples. While the sensitivity (for identifying S. aureus) remained satisfactory at 81%, the overall accuracy was compromised to 68% due to the substantial number of false positives. This could be attributed to the inherent limitations of the currently available smartphone-based microscope attachment, which makes focusing challenging, especially with skin swab samples. Overall, this work provides a proof-of-concept for identifying bacterial mixtures belonging to the same genus without using bioreceptors or laboratory equipment, toward low-cost, rapid assessment of complex bacterial mixture samples.