Fentanyl centered opioid overdose epidemic is afflicting the United States and worldwide. To address this crisis, point-of-care (POC) fentanyl testing is among critical health intervention tools, providing insights into its diffusion among populations. Deep learning-enabled image processing has significant applications in point-of-care diagnostics. However, the generalizability of traditional supervised convolutional neural networks may be precluded by the need for large, specialist-annotated training datasets.
We used a combination of smartphone-based bubbling microchips and catalytic platinum nanoparticles (PtNPs) to achieve high analytical sensitivity. We trained the network using a limited dataset of 104 images with known fentanyl concentrations and a retrospective library of 17,573 images, including 16,000 images synthetically generated through an adversarial neural network.
Fentanyl at concentrations as low as 0.23 ng/mL in phosphate buffered saline, 0.43 ng/mL in human serum and 0.64 ng/mL in artificial human urine can be detected. Classifier accuracy, using a cutoff concentration of 1 ng/mL, was 95.55% in PBS; 90.5% in human serum; and 90.63% in artificial human urine.
We developed the first deep learning assisted POC fentanyl detection with high sensitivity and accuracy.